Cog/ThirdParty/flac/patches/libflac-lzcnt.patch

diff --git a/src/libFLAC/bitreader.c b/src/libFLAC/bitreader.c
index 79cb5cc4..3ec9b90d 100644
--- a/src/libFLAC/bitreader.c
+++ b/src/libFLAC/bitreader.c
@@ -1098,3 +1098,238 @@ extern FLAC__bool FLAC__bitreader_is_consumed_byte_aligned(const FLAC__BitReader
 extern uint32_t FLAC__bitreader_bits_left_for_byte_alignment(const FLAC__BitReader *br);
 extern uint32_t FLAC__bitreader_get_input_bits_unconsumed(const FLAC__BitReader *br);
 extern FLAC__bool FLAC__bitreader_read_uint32_little_endian(FLAC__BitReader *br, FLAC__uint32 *val);
+
+
+
+#ifdef FLAC__SUPPORT_LZCNT
+
+/*
+* === LZCNT ===
+*/
+#if defined __GNUC__ || defined __clang__
+#define __LZCNT__
+#include <x86intrin.h>
+#endif
+
+#undef COUNT_ZERO_MSBS
+#undef COUNT_ZERO_MSBS2
+
+#if (ENABLE_64_BIT_WORDS == 0)
+#ifdef _MSC_VER
+#define COUNT_ZERO_MSBS(word) __lzcnt(word)
+#define COUNT_ZERO_MSBS2(word) __lzcnt(word)
+#else
+#define COUNT_ZERO_MSBS(word) __lzcnt32(word)
+#define COUNT_ZERO_MSBS2(word) __lzcnt32(word)
+#endif
+#else
+#define COUNT_ZERO_MSBS(word) __lzcnt64(word)
+#define COUNT_ZERO_MSBS2(word) __lzcnt64(word)
+#endif
+
+
+#ifdef __clang__
+#pragma clang attribute push (__attribute__((target("lzcnt"))), apply_to=function)
+#endif
+FLAC__bool FLAC__bitreader_read_unary_unsigned__LZCNT(FLAC__BitReader* br, uint32_t* val)
+{
+	uint32_t i;
+
+	FLAC__ASSERT(0 != br);
+	FLAC__ASSERT(0 != br->buffer);
+
+	*val = 0;
+	while(1) {
+		while(br->consumed_words < br->words) { /* if we've not consumed up to a partial tail word... */
+			brword b = br->buffer[br->consumed_words] << br->consumed_bits;
+			if(b) {
+				i = COUNT_ZERO_MSBS(b);
+				*val += i;
+				i++;
+				br->consumed_bits += i;
+				if (br->consumed_bits >= FLAC__BITS_PER_WORD) { /* faster way of testing if(br->consumed_bits == FLAC__BITS_PER_WORD) */
+					br->consumed_words++;
+					br->consumed_bits = 0;
+				}
+				return true;
+			} else {
+				*val += FLAC__BITS_PER_WORD - br->consumed_bits;
+				br->consumed_words++;
+				br->consumed_bits = 0;
+				/* didn't find stop bit yet, have to keep going... */
+			}
+		}
+		/* at this point we've eaten up all the whole words; have to try
+		 * reading through any tail bytes before calling the read callback.
+		 * this is a repeat of the above logic adjusted for the fact we
+		 * don't have a whole word.  note though if the client is feeding
+		 * us data a byte at a time (unlikely), br->consumed_bits may not
+		 * be zero.
+		 */
+		if(br->bytes * 8 > br->consumed_bits) {
+			const uint32_t end = br->bytes * 8;
+			brword b = (br->buffer[br->consumed_words] & (FLAC__WORD_ALL_ONES << (FLAC__BITS_PER_WORD - end))) << br->consumed_bits;
+			if(b) {
+				i = COUNT_ZERO_MSBS(b);
+				*val += i;
+				i++;
+				br->consumed_bits += i;
+				FLAC__ASSERT(br->consumed_bits < FLAC__BITS_PER_WORD);
+				return true;
+			} else {
+				*val += end - br->consumed_bits;
+				br->consumed_bits = end;
+				FLAC__ASSERT(br->consumed_bits < FLAC__BITS_PER_WORD);
+				/* didn't find stop bit yet, have to keep going... */
+			}
+		}
+		if(!bitreader_read_from_client_(br))
+			return false;
+	}
+}
+
+
+/* this is by far the most heavily used reader call.  it ain't pretty but it's fast */
+FLAC__bool FLAC__bitreader_read_rice_signed_block__LZCNT(FLAC__BitReader* br, int vals[], uint32_t nvals, uint32_t parameter)
+{
+	/* try and get br->consumed_words and br->consumed_bits into register;
+	 * must remember to flush them back to *br before calling other
+	 * bitreader functions that use them, and before returning */
+	uint32_t cwords, words, lsbs, msbs, x, y;
+	uint32_t ucbits; /* keep track of the number of unconsumed bits in word */
+	brword b;
+	int* val, * end;
+
+	FLAC__ASSERT(0 != br);
+	FLAC__ASSERT(0 != br->buffer);
+	/* WATCHOUT: code does not work with <32bit words; we can make things much faster with this assertion */
+	FLAC__ASSERT(FLAC__BITS_PER_WORD >= 32);
+	FLAC__ASSERT(parameter < 32);
+	/* the above two asserts also guarantee that the binary part never straddles more than 2 words, so we don't have to loop to read it */
+
+	val = vals;
+	end = vals + nvals;
+
+	if(parameter == 0) {
+		while(val < end) {
+			/* read the unary MSBs and end bit */
+			if(!FLAC__bitreader_read_unary_unsigned__LZCNT(br, &msbs))
+				return false;
+
+			*val++ = (int)(msbs >> 1) ^ -(int)(msbs & 1);
+		}
+
+		return true;
+	}
+
+	FLAC__ASSERT(parameter > 0);
+
+	cwords = br->consumed_words;
+	words = br->words;
+
+	/* if we've not consumed up to a partial tail word... */
+	if(cwords >= words) {
+		x = 0;
+		goto process_tail;
+	}
+
+	ucbits = FLAC__BITS_PER_WORD - br->consumed_bits;
+	b = br->buffer[cwords] << br->consumed_bits;  /* keep unconsumed bits aligned to left */
+
+	while(val < end) {
+		/* read the unary MSBs and end bit */
+		x = y = COUNT_ZERO_MSBS2(b);
+		if (x == FLAC__BITS_PER_WORD) {
+			x = ucbits;
+			do {
+				/* didn't find stop bit yet, have to keep going... */
+				cwords++;
+				if (cwords >= words)
+					goto incomplete_msbs;
+				b = br->buffer[cwords];
+				y = COUNT_ZERO_MSBS2(b);
+				x += y;
+			} while(y == FLAC__BITS_PER_WORD);
+		}
+		b <<= y;
+		b <<= 1; /* account for stop bit */
+		ucbits = (ucbits - x - 1) % FLAC__BITS_PER_WORD;
+		msbs = x;
+
+		/* read the binary LSBs */
+		x = (FLAC__uint32)(b >> (FLAC__BITS_PER_WORD - parameter)); /* parameter < 32, so we can cast to 32-bit uint32_t */
+		if(parameter <= ucbits) {
+			ucbits -= parameter;
+			b <<= parameter;
+		} else {
+			/* there are still bits left to read, they will all be in the next word */
+			cwords++;
+			if (cwords >= words)
+				goto incomplete_lsbs;
+			b = br->buffer[cwords];
+			ucbits += FLAC__BITS_PER_WORD - parameter;
+			x |= (FLAC__uint32)(b >> ucbits);
+			b <<= FLAC__BITS_PER_WORD - ucbits;
+		}
+		lsbs = x;
+
+		/* compose the value */
+		x = (msbs << parameter) | lsbs;
+		*val++ = (int)(x >> 1) ^ -(int)(x & 1);
+
+		continue;
+
+		/* at this point we've eaten up all the whole words */
+	process_tail:
+		do {
+			if(0) {
+			incomplete_msbs:
+				br->consumed_bits = 0;
+				br->consumed_words = cwords;
+			}
+
+			/* read the unary MSBs and end bit */
+			if(!FLAC__bitreader_read_unary_unsigned__LZCNT(br, &msbs))
+				return false;
+			msbs += x;
+			x = ucbits = 0;
+
+			if(0) {
+			incomplete_lsbs:
+				br->consumed_bits = 0;
+				br->consumed_words = cwords;
+			}
+
+			/* read the binary LSBs */
+			if(!FLAC__bitreader_read_raw_uint32(br, &lsbs, parameter - ucbits))
+				return false;
+			lsbs = x | lsbs;
+
+			/* compose the value */
+			x = (msbs << parameter) | lsbs;
+			*val++ = (int)(x >> 1) ^ -(int)(x & 1);
+			x = 0;
+
+			cwords = br->consumed_words;
+			words = br->words;
+			ucbits = FLAC__BITS_PER_WORD - br->consumed_bits;
+			b = br->buffer[cwords] << br->consumed_bits;
+		} while(cwords >= words && val < end);
+	}
+
+	if(ucbits == 0 && cwords < words) {
+		/* don't leave the head word with no unconsumed bits */
+		cwords++;
+		ucbits = FLAC__BITS_PER_WORD;
+	}
+
+	br->consumed_bits = FLAC__BITS_PER_WORD - ucbits;
+	br->consumed_words = cwords;
+
+	return true;
+}
+#ifdef __clang__
+#pragma clang attribute pop
+#endif
+
+#endif // FLAC__SUPPORT_LZCNT
\ No newline at end of file
diff --git a/src/libFLAC/cpu.c b/src/libFLAC/cpu.c
index 8b92f4c7..c0924512 100644
--- a/src/libFLAC/cpu.c
+++ b/src/libFLAC/cpu.c
@@ -195,6 +195,9 @@ x86_cpu_info (FLAC__CPUInfo *info)
 		info->x86.avx2  = (flags_ebx & FLAC__CPUINFO_X86_CPUID_AVX2   ) ? true : false;
 	}
 
+	cpuinfo_x86(0x80000001, &flags_eax, &flags_ebx, &flags_ecx, &flags_edx);
+	info->x86.lzcnt = (flags_ecx & (1 << 5)) ? true : false;
+
 #if defined FLAC__CPU_IA32
 	dfprintf(stderr, "CPU info (IA-32):\n");
 #else
diff --git a/src/libFLAC/include/private/bitreader.h b/src/libFLAC/include/private/bitreader.h
index 585a5db2..af37eb69 100644
--- a/src/libFLAC/include/private/bitreader.h
+++ b/src/libFLAC/include/private/bitreader.h
@@ -89,3 +89,9 @@ FLAC__bool FLAC__bitreader_read_golomb_unsigned(FLAC__BitReader *br, uint32_t *v
 FLAC__bool FLAC__bitreader_read_utf8_uint32(FLAC__BitReader *br, FLAC__uint32 *val, FLAC__byte *raw, uint32_t *rawlen);
 FLAC__bool FLAC__bitreader_read_utf8_uint64(FLAC__BitReader *br, FLAC__uint64 *val, FLAC__byte *raw, uint32_t *rawlen);
 #endif
+
+#if (defined FLAC__CPU_IA32 || defined FLAC__CPU_X86_64) && FLAC__HAS_X86INTRIN
+#define FLAC__SUPPORT_LZCNT
+FLAC__bool FLAC__bitreader_read_unary_unsigned__LZCNT(FLAC__BitReader* br, uint32_t* val);
+FLAC__bool FLAC__bitreader_read_rice_signed_block__LZCNT(FLAC__BitReader* br, int vals[], uint32_t nvals, uint32_t parameter);
+#endif
\ No newline at end of file
diff --git a/src/libFLAC/include/private/cpu.h b/src/libFLAC/include/private/cpu.h
index 0b50839f..61b71a37 100644
--- a/src/libFLAC/include/private/cpu.h
+++ b/src/libFLAC/include/private/cpu.h
@@ -178,6 +178,7 @@ typedef struct {
 	FLAC__bool avx;
 	FLAC__bool avx2;
 	FLAC__bool fma;
+	FLAC__bool lzcnt;
 } FLAC__CPUInfo_x86;
 
 typedef struct {
diff --git a/src/libFLAC/stream_decoder.c b/src/libFLAC/stream_decoder.c
index 4380b486..b1c4ea5f 100644
--- a/src/libFLAC/stream_decoder.c
+++ b/src/libFLAC/stream_decoder.c
@@ -134,6 +134,9 @@ typedef struct FLAC__StreamDecoderPrivate {
 	void (*local_lpc_restore_signal_64bit)(const FLAC__int32 residual[], uint32_t data_len, const FLAC__int32 qlp_coeff[], uint32_t order, int lp_quantization, FLAC__int32 data[]);
 	/* for use when the signal is <= 16 bits-per-sample, or <= 15 bits-per-sample on a side channel (which requires 1 extra bit): */
 	void (*local_lpc_restore_signal_16bit)(const FLAC__int32 residual[], uint32_t data_len, const FLAC__int32 qlp_coeff[], uint32_t order, int lp_quantization, FLAC__int32 data[]);
+
+	FLAC__bool (*local_bitreader_read_rice_signed_block)(FLAC__BitReader* br, int vals[], uint32_t nvals, uint32_t parameter);
+
 	void *client_data;
 	FILE *file; /* only used if FLAC__stream_decoder_init_file()/FLAC__stream_decoder_init_file() called, else NULL */
 	FLAC__BitReader *input;
@@ -377,9 +380,18 @@ static FLAC__StreamDecoderInitStatus init_stream_internal_(
 	decoder->private_->local_lpc_restore_signal = FLAC__lpc_restore_signal;
 	decoder->private_->local_lpc_restore_signal_64bit = FLAC__lpc_restore_signal_wide;
 	decoder->private_->local_lpc_restore_signal_16bit = FLAC__lpc_restore_signal;
+	decoder->private_->local_bitreader_read_rice_signed_block = FLAC__bitreader_read_rice_signed_block;
+
 	/* now override with asm where appropriate */
 #ifndef FLAC__NO_ASM
 	if(decoder->private_->cpuinfo.use_asm) {
+
+#ifdef FLAC__SUPPORT_LZCNT
+		if(decoder->private_->cpuinfo.x86.lzcnt) {
+			decoder->private_->local_bitreader_read_rice_signed_block = FLAC__bitreader_read_rice_signed_block__LZCNT;
+		}
+#endif
+
 #ifdef FLAC__CPU_IA32
 		FLAC__ASSERT(decoder->private_->cpuinfo.type == FLAC__CPUINFO_TYPE_IA32);
 #ifdef FLAC__HAS_NASM
@@ -2805,7 +2817,7 @@ FLAC__bool read_residual_partitioned_rice_(FLAC__StreamDecoder *decoder, uint32_
 		if(rice_parameter < pesc) {
 			partitioned_rice_contents->raw_bits[partition] = 0;
 			u = (partition == 0) ? partition_samples - predictor_order : partition_samples;
-			if(!FLAC__bitreader_read_rice_signed_block(decoder->private_->input, residual + sample, u, rice_parameter))
+			if(!decoder->private_->local_bitreader_read_rice_signed_block(decoder->private_->input, residual + sample, u, rice_parameter))
 				return false; /* read_callback_ sets the state for us */
 			sample += u;
 		}