libFLAC: update to 1.3.1

1 files changed, 1043 insertions, 107 deletions

diff --git a/sys/src/cmd/audio/libFLAC/lpc.c b/sys/src/cmd/audio/libFLAC/lpc.c
index 275b707a3..7eff63e6d 100644
--- a/sys/src/cmd/audio/libFLAC/lpc.c
+++ b/sys/src/cmd/audio/libFLAC/lpc.c
@@ -1,5 +1,6 @@
 /* libFLAC - Free Lossless Audio Codec library
- * Copyright (C) 2000,2001,2002,2003,2004  Josh Coalson
+ * Copyright (C) 2000-2009  Josh Coalson
+ * Copyright (C) 2011-2014  Xiph.Org Foundation
  *
  * Redistribution and use in source and binary forms, with or without
  * modification, are permitted provided that the following conditions
@@ -29,19 +30,48 @@
  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  */
 
+#ifdef HAVE_CONFIG_H
+#  include <config.h>
+#endif
+
 #include <math.h>
+
 #include "FLAC/assert.h"
 #include "FLAC/format.h"
+#include "share/compat.h"
 #include "private/bitmath.h"
 #include "private/lpc.h"
+#include "private/macros.h"
 #if defined DEBUG || defined FLAC__OVERFLOW_DETECT || defined FLAC__OVERFLOW_DETECT_VERBOSE
 #include <stdio.h>
 #endif
 
-#ifndef M_LN2
-/* math.h in VC++ doesn't seem to have this (how Microsoft is that?) */
-#define M_LN2 0.69314718055994530942
+/* OPT: #undef'ing this may improve the speed on some architectures */
+#define FLAC__LPC_UNROLLED_FILTER_LOOPS
+
+#ifndef FLAC__INTEGER_ONLY_LIBRARY
+
+#if !defined(HAVE_LROUND)
+#if defined(_MSC_VER)
+#include <float.h>
+#define copysign _copysign
+#elif defined(__GNUC__)
+#define copysign __builtin_copysign
+#elif defined(Plan9)
+#define copysign(x,y) ((x > 0.0 && y < 0.0) || (x < 0.0 && y > 0.0)) ? -x : x
 #endif
+static inline long int lround(double x) {
+    return (long)(x + copysign (0.5, x));
+}
+/* If this fails, we are in the presence of a mid 90's compiler, move along... */
+#endif
+
+void FLAC__lpc_window_data(const FLAC__int32 in[], const FLAC__real window[], FLAC__real out[], unsigned data_len)
+{
+	unsigned i;
+	for(i = 0; i < data_len; i++)
+		out[i] = in[i] * window[i];
+}
 
 void FLAC__lpc_compute_autocorrelation(const FLAC__real data[], unsigned data_len, unsigned lag, FLAC__real autoc[])
 {
@@ -53,6 +83,13 @@ void FLAC__lpc_compute_autocorrelation(const FLAC__real data[], unsigned data_le
 	FLAC__ASSERT(lag > 0);
 	FLAC__ASSERT(lag <= data_len);
 
+	/*
+	 * Technically we should subtract the mean first like so:
+	 *   for(i = 0; i < data_len; i++)
+	 *     data[i] -= mean;
+	 * but it appears not to make enough of a difference to matter, and
+	 * most signals are already closely centered around zero
+	 */
 	while(lag--) {
 		for(i = lag, d = 0.0; i < data_len; i++)
 			d += data[i] * data[i - lag];
@@ -85,28 +122,29 @@ void FLAC__lpc_compute_autocorrelation(const FLAC__real data[], unsigned data_le
 	}
 }
 
-void FLAC__lpc_compute_lp_coefficients(const FLAC__real autoc[], unsigned max_order, FLAC__real lp_coeff[][FLAC__MAX_LPC_ORDER], FLAC__real error[])
+void FLAC__lpc_compute_lp_coefficients(const FLAC__real autoc[], unsigned *max_order, FLAC__real lp_coeff[][FLAC__MAX_LPC_ORDER], FLAC__double error[])
 {
 	unsigned i, j;
-	double r, err, ref[FLAC__MAX_LPC_ORDER], lpc[FLAC__MAX_LPC_ORDER];
+	FLAC__double r, err, lpc[FLAC__MAX_LPC_ORDER];
 
-	FLAC__ASSERT(0 < max_order);
-	FLAC__ASSERT(max_order <= FLAC__MAX_LPC_ORDER);
+	FLAC__ASSERT(0 != max_order);
+	FLAC__ASSERT(0 < *max_order);
+	FLAC__ASSERT(*max_order <= FLAC__MAX_LPC_ORDER);
 	FLAC__ASSERT(autoc[0] != 0.0);
 
 	err = autoc[0];
 
-	for(i = 0; i < max_order; i++) {
+	for(i = 0; i < *max_order; i++) {
 		/* Sum up this iteration's reflection coefficient. */
 		r = -autoc[i+1];
 		for(j = 0; j < i; j++)
 			r -= lpc[j] * autoc[i-j];
-		ref[i] = (r/=err);
+		r /= err;
 
 		/* Update LPC coefficients and total error. */
 		lpc[i]=r;
 		for(j = 0; j < (i>>1); j++) {
-			double tmp = lpc[j];
+			FLAC__double tmp = lpc[j];
 			lpc[j] += r * lpc[i-1-j];
 			lpc[i-1-j] += r * tmp;
 		}
@@ -118,17 +156,21 @@ void FLAC__lpc_compute_lp_coefficients(const FLAC__real autoc[], unsigned max_or
 		/* save this order */
 		for(j = 0; j <= i; j++)
 			lp_coeff[i][j] = (FLAC__real)(-lpc[j]); /* negate FIR filter coeff to get predictor coeff */
-		error[i] = (FLAC__real)err;
+		error[i] = err;
+
+		/* see SF bug https://sourceforge.net/p/flac/bugs/234/ */
+		if(err == 0.0) {
+			*max_order = i+1;
+			return;
+		}
 	}
 }
 
 int FLAC__lpc_quantize_coefficients(const FLAC__real lp_coeff[], unsigned order, unsigned precision, FLAC__int32 qlp_coeff[], int *shift)
 {
 	unsigned i;
-	double d, cmax = -1e32;
+	FLAC__double cmax;
 	FLAC__int32 qmax, qmin;
-	const int max_shiftlimit = (1 << (FLAC__SUBFRAME_LPC_QLP_SHIFT_LEN-1)) - 1;
-	const int min_shiftlimit = -max_shiftlimit - 1;
 
 	FLAC__ASSERT(precision > 0);
 	FLAC__ASSERT(precision >= FLAC__MIN_QLP_COEFF_PRECISION);
@@ -139,77 +181,96 @@ int FLAC__lpc_quantize_coefficients(const FLAC__real lp_coeff[], unsigned order,
 	qmin = -qmax;
 	qmax--;
 
+	/* calc cmax = max( |lp_coeff[i]| ) */
+	cmax = 0.0;
 	for(i = 0; i < order; i++) {
-		if(lp_coeff[i] == 0.0)
-			continue;
-		d = fabs(lp_coeff[i]);
+		const FLAC__double d = fabs(lp_coeff[i]);
 		if(d > cmax)
 			cmax = d;
 	}
-redo_it:
+
 	if(cmax <= 0.0) {
 		/* => coefficients are all 0, which means our constant-detect didn't work */
 		return 2;
 	}
 	else {
+		const int max_shiftlimit = (1 << (FLAC__SUBFRAME_LPC_QLP_SHIFT_LEN-1)) - 1;
+		const int min_shiftlimit = -max_shiftlimit - 1;
 		int log2cmax;
 
 		(void)frexp(cmax, &log2cmax);
 		log2cmax--;
 		*shift = (int)precision - log2cmax - 1;
 
-		if(*shift < min_shiftlimit || *shift > max_shiftlimit) {
-#if 0
-			/*@@@ this does not seem to help at all, but was not extensively tested either: */
-			if(*shift > max_shiftlimit)
-				*shift = max_shiftlimit;
-			else
-#endif
-				return 1;
-		}
+		if(*shift > max_shiftlimit)
+			*shift = max_shiftlimit;
+		else if(*shift < min_shiftlimit)
+			return 1;
 	}
 
 	if(*shift >= 0) {
+		FLAC__double error = 0.0;
+		FLAC__int32 q;
 		for(i = 0; i < order; i++) {
-			qlp_coeff[i] = (FLAC__int32)floor((double)lp_coeff[i] * (double)(1 << *shift));
+			error += lp_coeff[i] * (1 << *shift);
+			q = lround(error);
 
-			/* double-check the result */
-			if(qlp_coeff[i] > qmax || qlp_coeff[i] < qmin) {
 #ifdef FLAC__OVERFLOW_DETECT
-				fprintf(stderr,"FLAC__lpc_quantize_coefficients: compensating for overflow, qlp_coeff[%u]=%d, lp_coeff[%u]=%f, cmax=%f, precision=%u, shift=%d, q=%f, f(q)=%f\n", i, qlp_coeff[i], i, lp_coeff[i], cmax, precision, *shift, (double)lp_coeff[i] * (double)(1 << *shift), floor((double)lp_coeff[i] * (double)(1 << *shift)));
+			if(q > qmax+1) /* we expect q==qmax+1 occasionally due to rounding */
+				fprintf(stderr,"FLAC__lpc_quantize_coefficients: quantizer overflow: q>qmax %d>%d shift=%d cmax=%f precision=%u lpc[%u]=%f\n",q,qmax,*shift,cmax,precision+1,i,lp_coeff[i]);
+			else if(q < qmin)
+				fprintf(stderr,"FLAC__lpc_quantize_coefficients: quantizer overflow: q<qmin %d<%d shift=%d cmax=%f precision=%u lpc[%u]=%f\n",q,qmin,*shift,cmax,precision+1,i,lp_coeff[i]);
 #endif
-				cmax *= 2.0;
-				goto redo_it;
-			}
+			if(q > qmax)
+				q = qmax;
+			else if(q < qmin)
+				q = qmin;
+			error -= q;
+			qlp_coeff[i] = q;
 		}
 	}
-	else { /* (*shift < 0) */
+	/* negative shift is very rare but due to design flaw, negative shift is
+	 * a NOP in the decoder, so it must be handled specially by scaling down
+	 * coeffs
+	 */
+	else {
 		const int nshift = -(*shift);
+		FLAC__double error = 0.0;
+		FLAC__int32 q;
 #ifdef DEBUG
-		fprintf(stderr,"FLAC__lpc_quantize_coefficients: negative shift = %d\n", *shift);
+		fprintf(stderr,"FLAC__lpc_quantize_coefficients: negative shift=%d order=%u cmax=%f\n", *shift, order, cmax);
 #endif
 		for(i = 0; i < order; i++) {
-			qlp_coeff[i] = (FLAC__int32)floor((double)lp_coeff[i] / (double)(1 << nshift));
-
-			/* double-check the result */
-			if(qlp_coeff[i] > qmax || qlp_coeff[i] < qmin) {
+			error += lp_coeff[i] / (1 << nshift);
+			q = lround(error);
 #ifdef FLAC__OVERFLOW_DETECT
-				fprintf(stderr,"FLAC__lpc_quantize_coefficients: compensating for overflow, qlp_coeff[%u]=%d, lp_coeff[%u]=%f, cmax=%f, precision=%u, shift=%d, q=%f, f(q)=%f\n", i, qlp_coeff[i], i, lp_coeff[i], cmax, precision, *shift, (double)lp_coeff[i] / (double)(1 << nshift), floor((double)lp_coeff[i] / (double)(1 << nshift)));
+			if(q > qmax+1) /* we expect q==qmax+1 occasionally due to rounding */
+				fprintf(stderr,"FLAC__lpc_quantize_coefficients: quantizer overflow: q>qmax %d>%d shift=%d cmax=%f precision=%u lpc[%u]=%f\n",q,qmax,*shift,cmax,precision+1,i,lp_coeff[i]);
+			else if(q < qmin)
+				fprintf(stderr,"FLAC__lpc_quantize_coefficients: quantizer overflow: q<qmin %d<%d shift=%d cmax=%f precision=%u lpc[%u]=%f\n",q,qmin,*shift,cmax,precision+1,i,lp_coeff[i]);
 #endif
-				cmax *= 2.0;
-				goto redo_it;
-			}
+			if(q > qmax)
+				q = qmax;
+			else if(q < qmin)
+				q = qmin;
+			error -= q;
+			qlp_coeff[i] = q;
 		}
+		*shift = 0;
 	}
 
 	return 0;
 }
 
-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[])
+#if defined(_MSC_VER)
+// silence MSVC warnings about __restrict modifier
+#pragma warning ( disable : 4028 )
+#endif
+
+void FLAC__lpc_compute_residual_from_qlp_coefficients(const FLAC__int32 * flac_restrict data, unsigned data_len, const FLAC__int32 * flac_restrict qlp_coeff, unsigned order, int lp_quantization, FLAC__int32 * flac_restrict residual)
+#if defined(FLAC__OVERFLOW_DETECT) || !defined(FLAC__LPC_UNROLLED_FILTER_LOOPS)
 {
-#ifdef FLAC__OVERFLOW_DETECT
 	FLAC__int64 sumo;
-#endif
 	unsigned i, j;
 	FLAC__int32 sum;
 	const FLAC__int32 *history;
@@ -223,23 +284,13 @@ void FLAC__lpc_compute_residual_from_qlp_coefficients(const FLAC__int32 data[],
 	FLAC__ASSERT(order > 0);
 
 	for(i = 0; i < data_len; i++) {
-#ifdef FLAC__OVERFLOW_DETECT
 		sumo = 0;
-#endif
 		sum = 0;
 		history = data;
 		for(j = 0; j < order; j++) {
 			sum += qlp_coeff[j] * (*(--history));
-#ifdef FLAC__OVERFLOW_DETECT
 			sumo += (FLAC__int64)qlp_coeff[j] * (FLAC__int64)(*history);
-#if defined _MSC_VER
-			if(sumo > 2147483647I64 || sumo < -2147483648I64)
-				fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients: OVERFLOW, i=%u, j=%u, c=%d, d=%d, sumo=%I64d\n",i,j,qlp_coeff[j],*history,sumo);
-#else
-			if(sumo > 2147483647ll || sumo < -2147483648ll)
-				fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients: OVERFLOW, i=%u, j=%u, c=%d, d=%d, sumo=%lld\n",i,j,qlp_coeff[j],*history,sumo);
-#endif
-#endif
+				fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients: OVERFLOW, i=%u, j=%u, c=%d, d=%d, sumo=%" PRId64 "\n",i,j,qlp_coeff[j],*history,sumo);
 		}
 		*(residual++) = *(data++) - (sum >> lp_quantization);
 	}
@@ -253,8 +304,231 @@ void FLAC__lpc_compute_residual_from_qlp_coefficients(const FLAC__int32 data[],
 	}
 	*/
 }
+#else /* fully unrolled version for normal use */
+{
+	int i;
+	FLAC__int32 sum;
+
+	FLAC__ASSERT(order > 0);
+	FLAC__ASSERT(order <= 32);
+
+	/*
+	 * We do unique versions up to 12th order since that's the subset limit.
+	 * Also they are roughly ordered to match frequency of occurrence to
+	 * minimize branching.
+	 */
+	if(order <= 12) {
+		if(order > 8) {
+			if(order > 10) {
+				if(order == 12) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[11] * data[i-12];
+						sum += qlp_coeff[10] * data[i-11];
+						sum += qlp_coeff[9] * data[i-10];
+						sum += qlp_coeff[8] * data[i-9];
+						sum += qlp_coeff[7] * data[i-8];
+						sum += qlp_coeff[6] * data[i-7];
+						sum += qlp_coeff[5] * data[i-6];
+						sum += qlp_coeff[4] * data[i-5];
+						sum += qlp_coeff[3] * data[i-4];
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						residual[i] = data[i] - (sum >> lp_quantization);
+					}
+				}
+				else { /* order == 11 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[10] * data[i-11];
+						sum += qlp_coeff[9] * data[i-10];
+						sum += qlp_coeff[8] * data[i-9];
+						sum += qlp_coeff[7] * data[i-8];
+						sum += qlp_coeff[6] * data[i-7];
+						sum += qlp_coeff[5] * data[i-6];
+						sum += qlp_coeff[4] * data[i-5];
+						sum += qlp_coeff[3] * data[i-4];
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						residual[i] = data[i] - (sum >> lp_quantization);
+					}
+				}
+			}
+			else {
+				if(order == 10) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[9] * data[i-10];
+						sum += qlp_coeff[8] * data[i-9];
+						sum += qlp_coeff[7] * data[i-8];
+						sum += qlp_coeff[6] * data[i-7];
+						sum += qlp_coeff[5] * data[i-6];
+						sum += qlp_coeff[4] * data[i-5];
+						sum += qlp_coeff[3] * data[i-4];
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						residual[i] = data[i] - (sum >> lp_quantization);
+					}
+				}
+				else { /* order == 9 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[8] * data[i-9];
+						sum += qlp_coeff[7] * data[i-8];
+						sum += qlp_coeff[6] * data[i-7];
+						sum += qlp_coeff[5] * data[i-6];
+						sum += qlp_coeff[4] * data[i-5];
+						sum += qlp_coeff[3] * data[i-4];
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						residual[i] = data[i] - (sum >> lp_quantization);
+					}
+				}
+			}
+		}
+		else if(order > 4) {
+			if(order > 6) {
+				if(order == 8) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[7] * data[i-8];
+						sum += qlp_coeff[6] * data[i-7];
+						sum += qlp_coeff[5] * data[i-6];
+						sum += qlp_coeff[4] * data[i-5];
+						sum += qlp_coeff[3] * data[i-4];
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						residual[i] = data[i] - (sum >> lp_quantization);
+					}
+				}
+				else { /* order == 7 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[6] * data[i-7];
+						sum += qlp_coeff[5] * data[i-6];
+						sum += qlp_coeff[4] * data[i-5];
+						sum += qlp_coeff[3] * data[i-4];
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						residual[i] = data[i] - (sum >> lp_quantization);
+					}
+				}
+			}
+			else {
+				if(order == 6) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[5] * data[i-6];
+						sum += qlp_coeff[4] * data[i-5];
+						sum += qlp_coeff[3] * data[i-4];
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						residual[i] = data[i] - (sum >> lp_quantization);
+					}
+				}
+				else { /* order == 5 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[4] * data[i-5];
+						sum += qlp_coeff[3] * data[i-4];
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						residual[i] = data[i] - (sum >> lp_quantization);
+					}
+				}
+			}
+		}
+		else {
+			if(order > 2) {
+				if(order == 4) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[3] * data[i-4];
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						residual[i] = data[i] - (sum >> lp_quantization);
+					}
+				}
+				else { /* order == 3 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						residual[i] = data[i] - (sum >> lp_quantization);
+					}
+				}
+			}
+			else {
+				if(order == 2) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						residual[i] = data[i] - (sum >> lp_quantization);
+					}
+				}
+				else { /* order == 1 */
+					for(i = 0; i < (int)data_len; i++)
+						residual[i] = data[i] - ((qlp_coeff[0] * data[i-1]) >> lp_quantization);
+				}
+			}
+		}
+	}
+	else { /* order > 12 */
+		for(i = 0; i < (int)data_len; i++) {
+			sum = 0;
+			switch(order) {
+				case 32: sum += qlp_coeff[31] * data[i-32];
+				case 31: sum += qlp_coeff[30] * data[i-31];
+				case 30: sum += qlp_coeff[29] * data[i-30];
+				case 29: sum += qlp_coeff[28] * data[i-29];
+				case 28: sum += qlp_coeff[27] * data[i-28];
+				case 27: sum += qlp_coeff[26] * data[i-27];
+				case 26: sum += qlp_coeff[25] * data[i-26];
+				case 25: sum += qlp_coeff[24] * data[i-25];
+				case 24: sum += qlp_coeff[23] * data[i-24];
+				case 23: sum += qlp_coeff[22] * data[i-23];
+				case 22: sum += qlp_coeff[21] * data[i-22];
+				case 21: sum += qlp_coeff[20] * data[i-21];
+				case 20: sum += qlp_coeff[19] * data[i-20];
+				case 19: sum += qlp_coeff[18] * data[i-19];
+				case 18: sum += qlp_coeff[17] * data[i-18];
+				case 17: sum += qlp_coeff[16] * data[i-17];
+				case 16: sum += qlp_coeff[15] * data[i-16];
+				case 15: sum += qlp_coeff[14] * data[i-15];
+				case 14: sum += qlp_coeff[13] * data[i-14];
+				case 13: sum += qlp_coeff[12] * data[i-13];
+				         sum += qlp_coeff[11] * data[i-12];
+				         sum += qlp_coeff[10] * data[i-11];
+				         sum += qlp_coeff[ 9] * data[i-10];
+				         sum += qlp_coeff[ 8] * data[i- 9];
+				         sum += qlp_coeff[ 7] * data[i- 8];
+				         sum += qlp_coeff[ 6] * data[i- 7];
+				         sum += qlp_coeff[ 5] * data[i- 6];
+				         sum += qlp_coeff[ 4] * data[i- 5];
+				         sum += qlp_coeff[ 3] * data[i- 4];
+				         sum += qlp_coeff[ 2] * data[i- 3];
+				         sum += qlp_coeff[ 1] * data[i- 2];
+				         sum += qlp_coeff[ 0] * data[i- 1];
+			}
+			residual[i] = data[i] - (sum >> lp_quantization);
+		}
+	}
+}
+#endif
 
-void FLAC__lpc_compute_residual_from_qlp_coefficients_wide(const FLAC__int32 data[], unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 residual[])
+void FLAC__lpc_compute_residual_from_qlp_coefficients_wide(const FLAC__int32 * flac_restrict data, unsigned data_len, const FLAC__int32 * flac_restrict qlp_coeff, unsigned order, int lp_quantization, FLAC__int32 * flac_restrict residual)
+#if defined(FLAC__OVERFLOW_DETECT) || !defined(FLAC__LPC_UNROLLED_FILTER_LOOPS)
 {
 	unsigned i, j;
 	FLAC__int64 sum;
@@ -273,28 +547,249 @@ void FLAC__lpc_compute_residual_from_qlp_coefficients_wide(const FLAC__int32 dat
 		history = data;
 		for(j = 0; j < order; j++)
 			sum += (FLAC__int64)qlp_coeff[j] * (FLAC__int64)(*(--history));
-#ifdef FLAC__OVERFLOW_DETECT
 		if(FLAC__bitmath_silog2_wide(sum >> lp_quantization) > 32) {
-			fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients_wide: OVERFLOW, i=%u, sum=%lld\n", i, sum >> lp_quantization);
+			fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients_wide: OVERFLOW, i=%u, sum=%" PRId64 "\n", i, (sum >> lp_quantization));
 			break;
 		}
 		if(FLAC__bitmath_silog2_wide((FLAC__int64)(*data) - (sum >> lp_quantization)) > 32) {
-			fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients_wide: OVERFLOW, i=%u, data=%d, sum=%lld, residual=%lld\n", i, *data, sum >> lp_quantization, (FLAC__int64)(*data) - (sum >> lp_quantization));
+			fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients_wide: OVERFLOW, i=%u, data=%d, sum=%" PRId64 ", residual=%" PRId64 "\n", i, *data, (int64_t)(sum >> lp_quantization), ((FLAC__int64)(*data) - (sum >> lp_quantization)));
 			break;
 		}
-#endif
 		*(residual++) = *(data++) - (FLAC__int32)(sum >> lp_quantization);
 	}
 }
+#else /* fully unrolled version for normal use */
+{
+	int i;
+	FLAC__int64 sum;
+
+	FLAC__ASSERT(order > 0);
+	FLAC__ASSERT(order <= 32);
+
+	/*
+	 * We do unique versions up to 12th order since that's the subset limit.
+	 * Also they are roughly ordered to match frequency of occurrence to
+	 * minimize branching.
+	 */
+	if(order <= 12) {
+		if(order > 8) {
+			if(order > 10) {
+				if(order == 12) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[11] * (FLAC__int64)data[i-12];
+						sum += qlp_coeff[10] * (FLAC__int64)data[i-11];
+						sum += qlp_coeff[9] * (FLAC__int64)data[i-10];
+						sum += qlp_coeff[8] * (FLAC__int64)data[i-9];
+						sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
+						sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
+						sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
+						sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
+						sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+				else { /* order == 11 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[10] * (FLAC__int64)data[i-11];
+						sum += qlp_coeff[9] * (FLAC__int64)data[i-10];
+						sum += qlp_coeff[8] * (FLAC__int64)data[i-9];
+						sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
+						sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
+						sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
+						sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
+						sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+			}
+			else {
+				if(order == 10) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[9] * (FLAC__int64)data[i-10];
+						sum += qlp_coeff[8] * (FLAC__int64)data[i-9];
+						sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
+						sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
+						sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
+						sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
+						sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+				else { /* order == 9 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[8] * (FLAC__int64)data[i-9];
+						sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
+						sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
+						sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
+						sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
+						sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+			}
+		}
+		else if(order > 4) {
+			if(order > 6) {
+				if(order == 8) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
+						sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
+						sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
+						sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
+						sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+				else { /* order == 7 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
+						sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
+						sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
+						sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+			}
+			else {
+				if(order == 6) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
+						sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
+						sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+				else { /* order == 5 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
+						sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+			}
+		}
+		else {
+			if(order > 2) {
+				if(order == 4) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+				else { /* order == 3 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+			}
+			else {
+				if(order == 2) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+				else { /* order == 1 */
+					for(i = 0; i < (int)data_len; i++)
+						residual[i] = data[i] - (FLAC__int32)((qlp_coeff[0] * (FLAC__int64)data[i-1]) >> lp_quantization);
+				}
+			}
+		}
+	}
+	else { /* order > 12 */
+		for(i = 0; i < (int)data_len; i++) {
+			sum = 0;
+			switch(order) {
+				case 32: sum += qlp_coeff[31] * (FLAC__int64)data[i-32];
+				case 31: sum += qlp_coeff[30] * (FLAC__int64)data[i-31];
+				case 30: sum += qlp_coeff[29] * (FLAC__int64)data[i-30];
+				case 29: sum += qlp_coeff[28] * (FLAC__int64)data[i-29];
+				case 28: sum += qlp_coeff[27] * (FLAC__int64)data[i-28];
+				case 27: sum += qlp_coeff[26] * (FLAC__int64)data[i-27];
+				case 26: sum += qlp_coeff[25] * (FLAC__int64)data[i-26];
+				case 25: sum += qlp_coeff[24] * (FLAC__int64)data[i-25];
+				case 24: sum += qlp_coeff[23] * (FLAC__int64)data[i-24];
+				case 23: sum += qlp_coeff[22] * (FLAC__int64)data[i-23];
+				case 22: sum += qlp_coeff[21] * (FLAC__int64)data[i-22];
+				case 21: sum += qlp_coeff[20] * (FLAC__int64)data[i-21];
+				case 20: sum += qlp_coeff[19] * (FLAC__int64)data[i-20];
+				case 19: sum += qlp_coeff[18] * (FLAC__int64)data[i-19];
+				case 18: sum += qlp_coeff[17] * (FLAC__int64)data[i-18];
+				case 17: sum += qlp_coeff[16] * (FLAC__int64)data[i-17];
+				case 16: sum += qlp_coeff[15] * (FLAC__int64)data[i-16];
+				case 15: sum += qlp_coeff[14] * (FLAC__int64)data[i-15];
+				case 14: sum += qlp_coeff[13] * (FLAC__int64)data[i-14];
+				case 13: sum += qlp_coeff[12] * (FLAC__int64)data[i-13];
+				         sum += qlp_coeff[11] * (FLAC__int64)data[i-12];
+				         sum += qlp_coeff[10] * (FLAC__int64)data[i-11];
+				         sum += qlp_coeff[ 9] * (FLAC__int64)data[i-10];
+				         sum += qlp_coeff[ 8] * (FLAC__int64)data[i- 9];
+				         sum += qlp_coeff[ 7] * (FLAC__int64)data[i- 8];
+				         sum += qlp_coeff[ 6] * (FLAC__int64)data[i- 7];
+				         sum += qlp_coeff[ 5] * (FLAC__int64)data[i- 6];
+				         sum += qlp_coeff[ 4] * (FLAC__int64)data[i- 5];
+				         sum += qlp_coeff[ 3] * (FLAC__int64)data[i- 4];
+				         sum += qlp_coeff[ 2] * (FLAC__int64)data[i- 3];
+				         sum += qlp_coeff[ 1] * (FLAC__int64)data[i- 2];
+				         sum += qlp_coeff[ 0] * (FLAC__int64)data[i- 1];
+			}
+			residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
+		}
+	}
+}
+#endif
+
+#endif /* !defined FLAC__INTEGER_ONLY_LIBRARY */
 
-void FLAC__lpc_restore_signal(const FLAC__int32 residual[], unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 data[])
+void FLAC__lpc_restore_signal(const FLAC__int32 * flac_restrict residual, unsigned data_len, const FLAC__int32 * flac_restrict qlp_coeff, unsigned order, int lp_quantization, FLAC__int32 * flac_restrict data)
+#if defined(FLAC__OVERFLOW_DETECT) || !defined(FLAC__LPC_UNROLLED_FILTER_LOOPS)
 {
-#ifdef FLAC__OVERFLOW_DETECT
 	FLAC__int64 sumo;
-#endif
 	unsigned i, j;
 	FLAC__int32 sum;
-	const FLAC__int32 *history;
+	const FLAC__int32 *r = residual, *history;
 
 #ifdef FLAC__OVERFLOW_DETECT_VERBOSE
 	fprintf(stderr,"FLAC__lpc_restore_signal: data_len=%d, order=%u, lpq=%d",data_len,order,lp_quantization);
@@ -305,25 +800,16 @@ void FLAC__lpc_restore_signal(const FLAC__int32 residual[], unsigned data_len, c
 	FLAC__ASSERT(order > 0);
 
 	for(i = 0; i < data_len; i++) {
-#ifdef FLAC__OVERFLOW_DETECT
 		sumo = 0;
-#endif
 		sum = 0;
 		history = data;
 		for(j = 0; j < order; j++) {
 			sum += qlp_coeff[j] * (*(--history));
-#ifdef FLAC__OVERFLOW_DETECT
 			sumo += (FLAC__int64)qlp_coeff[j] * (FLAC__int64)(*history);
-#if defined _MSC_VER
-			if(sumo > 2147483647I64 || sumo < -2147483648I64)
-				fprintf(stderr,"FLAC__lpc_restore_signal: OVERFLOW, i=%u, j=%u, c=%d, d=%d, sumo=%I64d\n",i,j,qlp_coeff[j],*history,sumo);
-#else
 			if(sumo > 2147483647ll || sumo < -2147483648ll)
-				fprintf(stderr,"FLAC__lpc_restore_signal: OVERFLOW, i=%u, j=%u, c=%d, d=%d, sumo=%lld\n",i,j,qlp_coeff[j],*history,sumo);
-#endif
-#endif
+				fprintf(stderr,"FLAC__lpc_restore_signal: OVERFLOW, i=%u, j=%u, c=%d, d=%d, sumo=%" PRId64 "\n",i,j,qlp_coeff[j],*history,sumo);
 		}
-		*(data++) = *(residual++) + (sum >> lp_quantization);
+		*(data++) = *(r++) + (sum >> lp_quantization);
 	}
 
 	/* Here's a slower but clearer version:
@@ -335,12 +821,235 @@ void FLAC__lpc_restore_signal(const FLAC__int32 residual[], unsigned data_len, c
 	}
 	*/
 }
+#else /* fully unrolled version for normal use */
+{
+	int i;
+	FLAC__int32 sum;
+
+	FLAC__ASSERT(order > 0);
+	FLAC__ASSERT(order <= 32);
+
+	/*
+	 * We do unique versions up to 12th order since that's the subset limit.
+	 * Also they are roughly ordered to match frequency of occurrence to
+	 * minimize branching.
+	 */
+	if(order <= 12) {
+		if(order > 8) {
+			if(order > 10) {
+				if(order == 12) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[11] * data[i-12];
+						sum += qlp_coeff[10] * data[i-11];
+						sum += qlp_coeff[9] * data[i-10];
+						sum += qlp_coeff[8] * data[i-9];
+						sum += qlp_coeff[7] * data[i-8];
+						sum += qlp_coeff[6] * data[i-7];
+						sum += qlp_coeff[5] * data[i-6];
+						sum += qlp_coeff[4] * data[i-5];
+						sum += qlp_coeff[3] * data[i-4];
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						data[i] = residual[i] + (sum >> lp_quantization);
+					}
+				}
+				else { /* order == 11 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[10] * data[i-11];
+						sum += qlp_coeff[9] * data[i-10];
+						sum += qlp_coeff[8] * data[i-9];
+						sum += qlp_coeff[7] * data[i-8];
+						sum += qlp_coeff[6] * data[i-7];
+						sum += qlp_coeff[5] * data[i-6];
+						sum += qlp_coeff[4] * data[i-5];
+						sum += qlp_coeff[3] * data[i-4];
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						data[i] = residual[i] + (sum >> lp_quantization);
+					}
+				}
+			}
+			else {
+				if(order == 10) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[9] * data[i-10];
+						sum += qlp_coeff[8] * data[i-9];
+						sum += qlp_coeff[7] * data[i-8];
+						sum += qlp_coeff[6] * data[i-7];
+						sum += qlp_coeff[5] * data[i-6];
+						sum += qlp_coeff[4] * data[i-5];
+						sum += qlp_coeff[3] * data[i-4];
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						data[i] = residual[i] + (sum >> lp_quantization);
+					}
+				}
+				else { /* order == 9 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[8] * data[i-9];
+						sum += qlp_coeff[7] * data[i-8];
+						sum += qlp_coeff[6] * data[i-7];
+						sum += qlp_coeff[5] * data[i-6];
+						sum += qlp_coeff[4] * data[i-5];
+						sum += qlp_coeff[3] * data[i-4];
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						data[i] = residual[i] + (sum >> lp_quantization);
+					}
+				}
+			}
+		}
+		else if(order > 4) {
+			if(order > 6) {
+				if(order == 8) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[7] * data[i-8];
+						sum += qlp_coeff[6] * data[i-7];
+						sum += qlp_coeff[5] * data[i-6];
+						sum += qlp_coeff[4] * data[i-5];
+						sum += qlp_coeff[3] * data[i-4];
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						data[i] = residual[i] + (sum >> lp_quantization);
+					}
+				}
+				else { /* order == 7 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[6] * data[i-7];
+						sum += qlp_coeff[5] * data[i-6];
+						sum += qlp_coeff[4] * data[i-5];
+						sum += qlp_coeff[3] * data[i-4];
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						data[i] = residual[i] + (sum >> lp_quantization);
+					}
+				}
+			}
+			else {
+				if(order == 6) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[5] * data[i-6];
+						sum += qlp_coeff[4] * data[i-5];
+						sum += qlp_coeff[3] * data[i-4];
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						data[i] = residual[i] + (sum >> lp_quantization);
+					}
+				}
+				else { /* order == 5 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[4] * data[i-5];
+						sum += qlp_coeff[3] * data[i-4];
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						data[i] = residual[i] + (sum >> lp_quantization);
+					}
+				}
+			}
+		}
+		else {
+			if(order > 2) {
+				if(order == 4) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[3] * data[i-4];
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						data[i] = residual[i] + (sum >> lp_quantization);
+					}
+				}
+				else { /* order == 3 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[2] * data[i-3];
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						data[i] = residual[i] + (sum >> lp_quantization);
+					}
+				}
+			}
+			else {
+				if(order == 2) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[1] * data[i-2];
+						sum += qlp_coeff[0] * data[i-1];
+						data[i] = residual[i] + (sum >> lp_quantization);
+					}
+				}
+				else { /* order == 1 */
+					for(i = 0; i < (int)data_len; i++)
+						data[i] = residual[i] + ((qlp_coeff[0] * data[i-1]) >> lp_quantization);
+				}
+			}
+		}
+	}
+	else { /* order > 12 */
+		for(i = 0; i < (int)data_len; i++) {
+			sum = 0;
+			switch(order) {
+				case 32: sum += qlp_coeff[31] * data[i-32];
+				case 31: sum += qlp_coeff[30] * data[i-31];
+				case 30: sum += qlp_coeff[29] * data[i-30];
+				case 29: sum += qlp_coeff[28] * data[i-29];
+				case 28: sum += qlp_coeff[27] * data[i-28];
+				case 27: sum += qlp_coeff[26] * data[i-27];
+				case 26: sum += qlp_coeff[25] * data[i-26];
+				case 25: sum += qlp_coeff[24] * data[i-25];
+				case 24: sum += qlp_coeff[23] * data[i-24];
+				case 23: sum += qlp_coeff[22] * data[i-23];
+				case 22: sum += qlp_coeff[21] * data[i-22];
+				case 21: sum += qlp_coeff[20] * data[i-21];
+				case 20: sum += qlp_coeff[19] * data[i-20];
+				case 19: sum += qlp_coeff[18] * data[i-19];
+				case 18: sum += qlp_coeff[17] * data[i-18];
+				case 17: sum += qlp_coeff[16] * data[i-17];
+				case 16: sum += qlp_coeff[15] * data[i-16];
+				case 15: sum += qlp_coeff[14] * data[i-15];
+				case 14: sum += qlp_coeff[13] * data[i-14];
+				case 13: sum += qlp_coeff[12] * data[i-13];
+				         sum += qlp_coeff[11] * data[i-12];
+				         sum += qlp_coeff[10] * data[i-11];
+				         sum += qlp_coeff[ 9] * data[i-10];
+				         sum += qlp_coeff[ 8] * data[i- 9];
+				         sum += qlp_coeff[ 7] * data[i- 8];
+				         sum += qlp_coeff[ 6] * data[i- 7];
+				         sum += qlp_coeff[ 5] * data[i- 6];
+				         sum += qlp_coeff[ 4] * data[i- 5];
+				         sum += qlp_coeff[ 3] * data[i- 4];
+				         sum += qlp_coeff[ 2] * data[i- 3];
+				         sum += qlp_coeff[ 1] * data[i- 2];
+				         sum += qlp_coeff[ 0] * data[i- 1];
+			}
+			data[i] = residual[i] + (sum >> lp_quantization);
+		}
+	}
+}
+#endif
 
-void FLAC__lpc_restore_signal_wide(const FLAC__int32 residual[], unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 data[])
+void FLAC__lpc_restore_signal_wide(const FLAC__int32 * flac_restrict residual, unsigned data_len, const FLAC__int32 * flac_restrict qlp_coeff, unsigned order, int lp_quantization, FLAC__int32 * flac_restrict data)
+#if defined(FLAC__OVERFLOW_DETECT) || !defined(FLAC__LPC_UNROLLED_FILTER_LOOPS)
 {
 	unsigned i, j;
 	FLAC__int64 sum;
-	const FLAC__int32 *history;
+	const FLAC__int32 *r = residual, *history;
 
 #ifdef FLAC__OVERFLOW_DETECT_VERBOSE
 	fprintf(stderr,"FLAC__lpc_restore_signal_wide: data_len=%d, order=%u, lpq=%d",data_len,order,lp_quantization);
@@ -355,69 +1064,296 @@ void FLAC__lpc_restore_signal_wide(const FLAC__int32 residual[], unsigned data_l
 		history = data;
 		for(j = 0; j < order; j++)
 			sum += (FLAC__int64)qlp_coeff[j] * (FLAC__int64)(*(--history));
-#ifdef FLAC__OVERFLOW_DETECT
 		if(FLAC__bitmath_silog2_wide(sum >> lp_quantization) > 32) {
-			fprintf(stderr,"FLAC__lpc_restore_signal_wide: OVERFLOW, i=%u, sum=%lld\n", i, sum >> lp_quantization);
+			fprintf(stderr,"FLAC__lpc_restore_signal_wide: OVERFLOW, i=%u, sum=%" PRId64 "\n", i, (sum >> lp_quantization));
 			break;
 		}
-		if(FLAC__bitmath_silog2_wide((FLAC__int64)(*residual) + (sum >> lp_quantization)) > 32) {
-			fprintf(stderr,"FLAC__lpc_restore_signal_wide: OVERFLOW, i=%u, residual=%d, sum=%lld, data=%lld\n", i, *residual, sum >> lp_quantization, (FLAC__int64)(*residual) + (sum >> lp_quantization));
+		if(FLAC__bitmath_silog2_wide((FLAC__int64)(*r) + (sum >> lp_quantization)) > 32) {
+			fprintf(stderr,"FLAC__lpc_restore_signal_wide: OVERFLOW, i=%u, residual=%d, sum=%" PRId64 ", data=%" PRId64 "\n", i, *r, (sum >> lp_quantization), ((FLAC__int64)(*r) + (sum >> lp_quantization)));
 			break;
 		}
-#endif
-		*(data++) = *(residual++) + (FLAC__int32)(sum >> lp_quantization);
+		*(data++) = *(r++) + (FLAC__int32)(sum >> lp_quantization);
 	}
 }
+#else /* fully unrolled version for normal use */
+{
+	int i;
+	FLAC__int64 sum;
 
-FLAC__real FLAC__lpc_compute_expected_bits_per_residual_sample(FLAC__real lpc_error, unsigned total_samples)
+	FLAC__ASSERT(order > 0);
+	FLAC__ASSERT(order <= 32);
+
+	/*
+	 * We do unique versions up to 12th order since that's the subset limit.
+	 * Also they are roughly ordered to match frequency of occurrence to
+	 * minimize branching.
+	 */
+	if(order <= 12) {
+		if(order > 8) {
+			if(order > 10) {
+				if(order == 12) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[11] * (FLAC__int64)data[i-12];
+						sum += qlp_coeff[10] * (FLAC__int64)data[i-11];
+						sum += qlp_coeff[9] * (FLAC__int64)data[i-10];
+						sum += qlp_coeff[8] * (FLAC__int64)data[i-9];
+						sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
+						sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
+						sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
+						sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
+						sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+				else { /* order == 11 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[10] * (FLAC__int64)data[i-11];
+						sum += qlp_coeff[9] * (FLAC__int64)data[i-10];
+						sum += qlp_coeff[8] * (FLAC__int64)data[i-9];
+						sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
+						sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
+						sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
+						sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
+						sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+			}
+			else {
+				if(order == 10) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[9] * (FLAC__int64)data[i-10];
+						sum += qlp_coeff[8] * (FLAC__int64)data[i-9];
+						sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
+						sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
+						sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
+						sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
+						sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+				else { /* order == 9 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[8] * (FLAC__int64)data[i-9];
+						sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
+						sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
+						sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
+						sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
+						sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+			}
+		}
+		else if(order > 4) {
+			if(order > 6) {
+				if(order == 8) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
+						sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
+						sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
+						sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
+						sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+				else { /* order == 7 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
+						sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
+						sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
+						sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+			}
+			else {
+				if(order == 6) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
+						sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
+						sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+				else { /* order == 5 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
+						sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+			}
+		}
+		else {
+			if(order > 2) {
+				if(order == 4) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+				else { /* order == 3 */
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+			}
+			else {
+				if(order == 2) {
+					for(i = 0; i < (int)data_len; i++) {
+						sum = 0;
+						sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
+						sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
+						data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
+					}
+				}
+				else { /* order == 1 */
+					for(i = 0; i < (int)data_len; i++)
+						data[i] = residual[i] + (FLAC__int32)((qlp_coeff[0] * (FLAC__int64)data[i-1]) >> lp_quantization);
+				}
+			}
+		}
+	}
+	else { /* order > 12 */
+		for(i = 0; i < (int)data_len; i++) {
+			sum = 0;
+			switch(order) {
+				case 32: sum += qlp_coeff[31] * (FLAC__int64)data[i-32];
+				case 31: sum += qlp_coeff[30] * (FLAC__int64)data[i-31];
+				case 30: sum += qlp_coeff[29] * (FLAC__int64)data[i-30];
+				case 29: sum += qlp_coeff[28] * (FLAC__int64)data[i-29];
+				case 28: sum += qlp_coeff[27] * (FLAC__int64)data[i-28];
+				case 27: sum += qlp_coeff[26] * (FLAC__int64)data[i-27];
+				case 26: sum += qlp_coeff[25] * (FLAC__int64)data[i-26];
+				case 25: sum += qlp_coeff[24] * (FLAC__int64)data[i-25];
+				case 24: sum += qlp_coeff[23] * (FLAC__int64)data[i-24];
+				case 23: sum += qlp_coeff[22] * (FLAC__int64)data[i-23];
+				case 22: sum += qlp_coeff[21] * (FLAC__int64)data[i-22];
+				case 21: sum += qlp_coeff[20] * (FLAC__int64)data[i-21];
+				case 20: sum += qlp_coeff[19] * (FLAC__int64)data[i-20];
+				case 19: sum += qlp_coeff[18] * (FLAC__int64)data[i-19];
+				case 18: sum += qlp_coeff[17] * (FLAC__int64)data[i-18];
+				case 17: sum += qlp_coeff[16] * (FLAC__int64)data[i-17];
+				case 16: sum += qlp_coeff[15] * (FLAC__int64)data[i-16];
+				case 15: sum += qlp_coeff[14] * (FLAC__int64)data[i-15];
+				case 14: sum += qlp_coeff[13] * (FLAC__int64)data[i-14];
+				case 13: sum += qlp_coeff[12] * (FLAC__int64)data[i-13];
+				         sum += qlp_coeff[11] * (FLAC__int64)data[i-12];
+				         sum += qlp_coeff[10] * (FLAC__int64)data[i-11];
+				         sum += qlp_coeff[ 9] * (FLAC__int64)data[i-10];
+				         sum += qlp_coeff[ 8] * (FLAC__int64)data[i- 9];
+				         sum += qlp_coeff[ 7] * (FLAC__int64)data[i- 8];
+				         sum += qlp_coeff[ 6] * (FLAC__int64)data[i- 7];
+				         sum += qlp_coeff[ 5] * (FLAC__int64)data[i- 6];
+				         sum += qlp_coeff[ 4] * (FLAC__int64)data[i- 5];
+				         sum += qlp_coeff[ 3] * (FLAC__int64)data[i- 4];
+				         sum += qlp_coeff[ 2] * (FLAC__int64)data[i- 3];
+				         sum += qlp_coeff[ 1] * (FLAC__int64)data[i- 2];
+				         sum += qlp_coeff[ 0] * (FLAC__int64)data[i- 1];
+			}
+			data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
+		}
+	}
+}
+#endif
+
+#if defined(_MSC_VER)
+#pragma warning ( default : 4028 )
+#endif
+
+#ifndef FLAC__INTEGER_ONLY_LIBRARY
+
+FLAC__double FLAC__lpc_compute_expected_bits_per_residual_sample(FLAC__double lpc_error, unsigned total_samples)
 {
-	double error_scale;
+	FLAC__double error_scale;
 
 	FLAC__ASSERT(total_samples > 0);
 
-	error_scale = 0.5 * M_LN2 * M_LN2 / (FLAC__real)total_samples;
+	error_scale = 0.5 * M_LN2 * M_LN2 / (FLAC__double)total_samples;
 
 	return FLAC__lpc_compute_expected_bits_per_residual_sample_with_error_scale(lpc_error, error_scale);
 }
 
-FLAC__real FLAC__lpc_compute_expected_bits_per_residual_sample_with_error_scale(FLAC__real lpc_error, double error_scale)
+FLAC__double FLAC__lpc_compute_expected_bits_per_residual_sample_with_error_scale(FLAC__double lpc_error, FLAC__double error_scale)
 {
 	if(lpc_error > 0.0) {
-		FLAC__real bps = (FLAC__real)((double)0.5 * log(error_scale * lpc_error) / M_LN2);
+		FLAC__double bps = (FLAC__double)0.5 * log(error_scale * lpc_error) / M_LN2;
 		if(bps >= 0.0)
 			return bps;
 		else
 			return 0.0;
 	}
-	else if(lpc_error < 0.0) { /* error should not be negative but can happen due to inadequate float resolution */
-		return (FLAC__real)1e32;
+	else if(lpc_error < 0.0) { /* error should not be negative but can happen due to inadequate floating-point resolution */
+		return 1e32;
 	}
 	else {
 		return 0.0;
 	}
 }
 
-unsigned FLAC__lpc_compute_best_order(const FLAC__real lpc_error[], unsigned max_order, unsigned total_samples, unsigned bits_per_signal_sample)
+unsigned FLAC__lpc_compute_best_order(const FLAC__double lpc_error[], unsigned max_order, unsigned total_samples, unsigned overhead_bits_per_order)
 {
-	unsigned order, best_order;
-	FLAC__real best_bits, tmp_bits;
-	double error_scale;
+	unsigned order, indx, best_index; /* 'index' the index into lpc_error; index==order-1 since lpc_error[0] is for order==1, lpc_error[1] is for order==2, etc */
+	FLAC__double bits, best_bits, error_scale;
 
 	FLAC__ASSERT(max_order > 0);
 	FLAC__ASSERT(total_samples > 0);
 
-	error_scale = 0.5 * M_LN2 * M_LN2 / (FLAC__real)total_samples;
+	error_scale = 0.5 * M_LN2 * M_LN2 / (FLAC__double)total_samples;
 
-	best_order = 0;
-	best_bits = FLAC__lpc_compute_expected_bits_per_residual_sample_with_error_scale(lpc_error[0], error_scale) * (FLAC__real)total_samples;
+	best_index = 0;
+	best_bits = (unsigned)(-1);
 
-	for(order = 1; order < max_order; order++) {
-		tmp_bits = FLAC__lpc_compute_expected_bits_per_residual_sample_with_error_scale(lpc_error[order], error_scale) * (FLAC__real)(total_samples - order) + (FLAC__real)(order * bits_per_signal_sample);
-		if(tmp_bits < best_bits) {
-			best_order = order;
-			best_bits = tmp_bits;
+	for(indx = 0, order = 1; indx < max_order; indx++, order++) {
+		bits = FLAC__lpc_compute_expected_bits_per_residual_sample_with_error_scale(lpc_error[indx], error_scale) * (FLAC__double)(total_samples - order) + (FLAC__double)(order * overhead_bits_per_order);
+		if(bits < best_bits) {
+			best_index = indx;
+			best_bits = bits;
 		}
 	}
 
-	return best_order+1; /* +1 since index of lpc_error[] is order-1 */
+	return best_index+1; /* +1 since indx of lpc_error[] is order-1 */
 }
+
+#endif /* !defined FLAC__INTEGER_ONLY_LIBRARY */