From 56361b52e7de7e235b72bd8241589537a1a40b18 Mon Sep 17 00:00:00 2001
From: Alibek Omarov <a1ba.omarov@gmail.com>
Date: Sat, 2 Aug 2025 12:18:18 +0500
Subject: [PATCH] 3rdparty: opus: enable platform optimized code (x86 SSE/AVX
 with runtime detection and NEON on ARM)

Along that enable hardening and check for C99 lrintf (probably can do the same in the engine at some point).
---
 3rdparty/opus/wscript | 71 +++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 68 insertions(+), 3 deletions(-)

diff --git a/3rdparty/opus/wscript b/3rdparty/opus/wscript
index 211cb10b..ed473963 100644
--- a/3rdparty/opus/wscript
+++ b/3rdparty/opus/wscript
@@ -23,6 +23,17 @@ int main (void) {
 	int * array = alloca(foo);
 }'''
 
+FRAGMENT_LRINT='''#include <math.h>
+#include <stdlib.h>
+int main (int argc, char **argv) {
+	return lrint%s((%s)atof(argv[1]));
+}'''
+
+# assuming MSVC always enables NEON
+FRAGMENT_NEON='''#if !defined __ARM_NEON__ && !defined _MSC_VER
+#error
+#endif'''
+
 def options(opt):
 	pass
 
@@ -31,15 +42,59 @@ def configure(conf):
 		conf.fatal('Can\'t find opus submodule. Run `git submodule update --init --recursive`.')
 		return
 
+	if conf.check_cc(fragment=FRAGMENT_LRINT % ('', 'double'), msg = 'Checking for C99 lrint', use = 'M', mandatory = False):
+		conf.define('HAVE_LRINT', 1)
+
+	if conf.check_cc(fragment=FRAGMENT_LRINT % ('f', 'float'), msg = 'Checking for C99 lrintf', use = 'M', mandatory = False):
+		conf.define('HAVE_LRINTF', 1)
+
 	# Check for C99 variable-size arrays, or alloca() as fallback
 	if conf.check_cc(fragment=FRAGMENT_VLA, msg = 'Checking for C99 VLA support', mandatory = False):
 		conf.define('VAR_ARRAYS', 1)
 	elif conf.check_cc(fragment=FRAGMENT_ALLOCA_H, msg = 'Checking for alloca in alloca.h header', mandatory = False):
 		conf.define('USE_ALLOCA', 1)
 		conf.define('HAVE_ALLOCA_H', 1)
-	elif conf.check_cc(fragmenmt=FRAGMENT_STDLIB_H, msg = 'Checking for alloca.h in stdlib.h', mandatory = False):
+	elif conf.check_cc(fragment=FRAGMENT_STDLIB_H, msg = 'Checking for alloca.h in stdlib.h', mandatory = False):
 		conf.define('USE_ALLOCA', 1)
 
+	if conf.env.DEST_CPU in ['thumb', 'arm']:
+		if conf.check(header_name='arm_neon.h', mandatory = False):
+			if conf.check(fragment=FRAGMENT_NEON, msg = 'Checking if compiler enabled NEON', mandatory = False):
+				conf.env.HAVE_NEON = True
+				conf.define('OPUS_ARM_MAY_HAVE_NEON_INTR', 1)
+				conf.define('OPUS_ARM_PRESUME_NEON', 1)
+	elif conf.env.DEST_CPU.startswith('x86'):
+		if conf.check(header_name='cpuid.h', mandatory=False):
+			conf.define('CPU_INFO_BY_C', 1)
+		else:
+			conf.define('CPU_INFO_BY_ASM', 1)
+
+		conf.define('OPUS_HAVE_RTCD', 1)
+
+		if conf.check(header_name='xmmintrin.h', mandatory = False):
+			if conf.env.COMPILER_CC != 'msvc':
+				conf.env.CFLAGS += ['-msse']
+			conf.define('OPUS_X86_MAY_HAVE_SSE', 1)
+			if conf.env.DEST_SIZEOF_VOID_P > 4:
+				conf.define('OPUS_X86_PRESUME_SSE', 1)
+
+		if conf.check(header_name='emmintrin.h', mandatory = False):
+			if conf.env.COMPILER_CC != 'msvc':
+				conf.env.CFLAGS += ['-msse2']
+			conf.define('OPUS_X86_MAY_HAVE_SSE2', 1)
+			if conf.env.DEST_SIZEOF_VOID_P > 4:
+				conf.define('OPUS_X86_PRESUME_SSE2', 1)
+
+		if conf.check(header_name='smmintrin.h', mandatory = False):
+			if conf.env.COMPILER_CC != 'msvc':
+				conf.env.CFLAGS += ['-msse4.1']
+			conf.define('OPUS_X86_MAY_HAVE_SSE4_1', 1)
+
+		if conf.check(header_name='immintrin.h', mandatory = False):
+			if conf.env.COMPILER_CC != 'msvc':
+				conf.env.CFLAGS += ['-mavx']
+			conf.define('OPUS_X86_MAY_HAVE_AVX', 1)
+
 	# TODO: ARM/x86 intrinsics detection
 	# TODO: maybe call autotools/cmake/meson instead?
 
@@ -55,8 +110,18 @@ def build(bld):
 		'opus/src/opus_compare.c',
 		'opus/celt/opus_custom_demo.c'
 	])
-	includes = ['opus/include/', 'opus/celt/', 'opus/silk/', 'opus/silk/float/']
-	defines = ['OPUS_BUILD', 'FLOAT_APPROX', 'PACKAGE_VERSION="1.4.0"', 'CUSTOM_MODES']
+
+	includes = ['opus', 'opus/include/', 'opus/celt/', 'opus/silk/', 'opus/silk/float/']
+
+	if bld.env.DEST_CPU in ['thumb', 'arm']:
+		if bld.env.HAVE_NEON:
+			sources += bld.path.ant_glob(['opus/silk/arm/*.c', 'opus/celt/arm/*.c'])
+			includes += ['opus/silk/arm', 'opus/celt/arm']
+	elif bld.env.DEST_CPU.startswith('x86'):
+		sources += bld.path.ant_glob(['opus/silk/x86/*.c', 'opus/celt/x86/*.c'])
+		includes += ['opus/silk/x86', 'opus/celt/x86']
+
+	defines = ['OPUS_BUILD', 'FLOAT_APPROX', 'PACKAGE_VERSION="1.4.0"', 'CUSTOM_MODES', 'ENABLE_HARDENING']
 
 	bld.stlib(
 		source = sources,