diff --git a/src/libFLAC/stream_decoder.c b/src/libFLAC/stream_decoder.c
index 7ccc526..ddd8979 100644
--- a/src/libFLAC/stream_decoder.c
+++ b/src/libFLAC/stream_decoder.c
@@ -404,7 +404,7 @@ static FLAC__StreamDecoderInitStatus init_stream_internal_(
 		if(decoder->private_->cpuinfo.ia32.bswap)
 			decoder->private_->local_bitreader_read_rice_signed_block = FLAC__bitreader_read_rice_signed_block_asm_ia32_bswap;
 #endif
-		decoder->private_->local_lpc_restore_signal_64bit = FLAC__lpc_restore_signal_wide_asm_ia32;
+		decoder->private_->local_lpc_restore_signal_64bit = FLAC__lpc_restore_signal_wide_asm_ia32; /* OPT_IA32: was really necessary for GCC < 4.9 */
 		if(decoder->private_->cpuinfo.ia32.mmx) {
 			decoder->private_->local_lpc_restore_signal = FLAC__lpc_restore_signal_asm_ia32;
 			decoder->private_->local_lpc_restore_signal_16bit = FLAC__lpc_restore_signal_asm_ia32_mmx;
@@ -417,13 +417,13 @@ static FLAC__StreamDecoderInitStatus init_stream_internal_(
 		}
 #endif
 #ifdef FLAC__HAS_X86INTRIN
-# if defined FLAC__SSE2_SUPPORTED && !defined FLAC__HAS_NASM /* OPT: not faster than ASM/MMX code */
+# if defined FLAC__SSE2_SUPPORTED && !defined FLAC__HAS_NASM /* OPT_SSE: not faster than ASM/MMX code */
 		if(decoder->private_->cpuinfo.ia32.sse2) {
 			decoder->private_->local_lpc_restore_signal_16bit = FLAC__lpc_restore_signal_16_intrin_sse2;
 			decoder->private_->local_lpc_restore_signal_16bit_order8 = FLAC__lpc_restore_signal_16_intrin_sse2;
 		}
 # endif
-# if defined FLAC__SSE4_1_SUPPORTED && 1 /* OPT: faster than asm; TODO: more tests */
+# if defined FLAC__SSE4_1_SUPPORTED && 1 /* OPT_SSE: faster than asm; TODO: more tests */
 		if(decoder->private_->cpuinfo.ia32.sse41)
 			decoder->private_->local_lpc_restore_signal_64bit = FLAC__lpc_restore_signal_wide_intrin_sse41;
 # endif
diff --git a/src/libFLAC/stream_encoder.c b/src/libFLAC/stream_encoder.c
index e21f374..e64ece2 100644
--- a/src/libFLAC/stream_encoder.c
+++ b/src/libFLAC/stream_encoder.c
@@ -894,7 +894,6 @@ static FLAC__StreamEncoderInitStatus init_stream_internal_(
 #  ifdef FLAC__CPU_IA32
 		FLAC__ASSERT(encoder->private_->cpuinfo.type == FLAC__CPUINFO_TYPE_IA32);
 #   ifdef FLAC__HAS_NASM
-		encoder->private_->local_lpc_compute_residual_from_qlp_coefficients_64bit = FLAC__lpc_compute_residual_from_qlp_coefficients_wide_asm_ia32;
 		if(encoder->private_->cpuinfo.ia32.sse) {
 			if(encoder->protected_->max_lpc_order < 4)
 				encoder->private_->local_lpc_compute_autocorrelation = FLAC__lpc_compute_autocorrelation_asm_ia32_sse_lag_4;
@@ -908,9 +907,11 @@ static FLAC__StreamEncoderInitStatus init_stream_internal_(
 				encoder->private_->local_lpc_compute_autocorrelation = FLAC__lpc_compute_autocorrelation_asm_ia32;
 		}
 		else if(encoder->private_->cpuinfo.ia32._3dnow)
-			encoder->private_->local_lpc_compute_autocorrelation = FLAC__lpc_compute_autocorrelation_asm_ia32_3dnow;
+			encoder->private_->local_lpc_compute_autocorrelation = FLAC__lpc_compute_autocorrelation_asm_ia32_3dnow; /* obsolete instruction set */
 		else
 			encoder->private_->local_lpc_compute_autocorrelation = FLAC__lpc_compute_autocorrelation_asm_ia32;
+
+		encoder->private_->local_lpc_compute_residual_from_qlp_coefficients_64bit = FLAC__lpc_compute_residual_from_qlp_coefficients_wide_asm_ia32; /* OPT_IA32: was really necessary for GCC < 4.9 */
 		if(encoder->private_->cpuinfo.ia32.mmx) {
 			encoder->private_->local_lpc_compute_residual_from_qlp_coefficients = FLAC__lpc_compute_residual_from_qlp_coefficients_asm_ia32;
 			encoder->private_->local_lpc_compute_residual_from_qlp_coefficients_16bit = FLAC__lpc_compute_residual_from_qlp_coefficients_asm_ia32_mmx;
@@ -919,11 +920,12 @@ static FLAC__StreamEncoderInitStatus init_stream_internal_(
 			encoder->private_->local_lpc_compute_residual_from_qlp_coefficients = FLAC__lpc_compute_residual_from_qlp_coefficients_asm_ia32;
 			encoder->private_->local_lpc_compute_residual_from_qlp_coefficients_16bit = FLAC__lpc_compute_residual_from_qlp_coefficients_asm_ia32;
 		}
+
 		if(encoder->private_->cpuinfo.ia32.mmx && encoder->private_->cpuinfo.ia32.cmov)
 			encoder->private_->local_fixed_compute_best_predictor = FLAC__fixed_compute_best_predictor_asm_ia32_mmx_cmov;
 #   endif /* FLAC__HAS_NASM */
 #   ifdef FLAC__HAS_X86INTRIN
-#    if defined FLAC__SSE_SUPPORTED && !defined FLAC__HAS_NASM
+#    if defined FLAC__SSE_SUPPORTED
 		if(encoder->private_->cpuinfo.ia32.sse) {
 			if(encoder->protected_->max_lpc_order < 4)
 				encoder->private_->local_lpc_compute_autocorrelation = FLAC__lpc_compute_autocorrelation_intrin_sse_lag_4;
@@ -940,6 +942,12 @@ static FLAC__StreamEncoderInitStatus init_stream_internal_(
 			encoder->private_->local_lpc_compute_residual_from_qlp_coefficients = FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse2;
 			encoder->private_->local_lpc_compute_residual_from_qlp_coefficients_16bit = FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2;
 		}
+#     ifdef FLAC__SSE4_1_SUPPORTED
+		if(encoder->private_->cpuinfo.ia32.sse41) {
+			encoder->private_->local_lpc_compute_residual_from_qlp_coefficients_64bit = FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_sse41;
+		}
+#     endif
+
 #     ifdef FLAC__SSSE3_SUPPORTED
 		if (encoder->private_->cpuinfo.ia32.ssse3) {
 			encoder->private_->local_fixed_compute_best_predictor = FLAC__fixed_compute_best_predictor_intrin_ssse3;
@@ -952,10 +960,6 @@ static FLAC__StreamEncoderInitStatus init_stream_internal_(
 			encoder->private_->local_fixed_compute_best_predictor_wide = FLAC__fixed_compute_best_predictor_wide_intrin_sse2;
 		}
 #    endif
-#    ifdef FLAC__SSE4_1_SUPPORTED
-		if(encoder->private_->cpuinfo.ia32.sse41)
-			encoder->private_->local_lpc_compute_residual_from_qlp_coefficients_64bit = FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_sse41;
-#    endif
 #   endif /* FLAC__HAS_X86INTRIN */
 #  elif defined FLAC__CPU_X86_64
 		FLAC__ASSERT(encoder->private_->cpuinfo.type == FLAC__CPUINFO_TYPE_X86_64);
@@ -971,8 +975,9 @@ static FLAC__StreamEncoderInitStatus init_stream_internal_(
 			encoder->private_->local_lpc_compute_autocorrelation = FLAC__lpc_compute_autocorrelation_intrin_sse_lag_16;
 #    endif
 #    ifdef FLAC__SSE2_SUPPORTED
-		/* encoder->private_->local_lpc_compute_residual_from_qlp_coefficients = FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse2; // OPT: not faster than C; TODO: more tests on different CPUs */
+		/* encoder->private_->local_lpc_compute_residual_from_qlp_coefficients = FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse2; // OPT_SSE: not faster than C; TODO: more tests on different CPUs */
 		encoder->private_->local_lpc_compute_residual_from_qlp_coefficients_16bit = FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_sse2;
+
 #     ifdef FLAC__SSSE3_SUPPORTED
 		if (encoder->private_->cpuinfo.x86_64.ssse3) {
 			encoder->private_->local_fixed_compute_best_predictor = FLAC__fixed_compute_best_predictor_intrin_ssse3;