commit 671f2daa29ba5190ef68a7afb2d907f2ea08db7c
author Behdad Esfahbod <behdad@behdad.org> Mon Oct 07 06:15:10 2024 -0600
committer Behdad Esfahbod <behdad@behdad.org> Wed Oct 09 10:31:49 2024 -0600
tree bec41dda77b38d1b3e941c4c3a080ef76392023a
parent d70ab4c979cbd143a273152aa2d78baafca40cde

[CONFIG] Add HB_NO_SIMD

src/hb-config.hh

diff --git a/src/hb-config.hh b/src/hb-config.hh
index 1410584..ad3c127 100644
--- a/src/hb-config.hh
+++ b/src/hb-config.hh
@@ -48,6 +48,7 @@
 #define HB_OPTIMIZE_SIZE_MORE
 #define HB_MINIMIZE_MEMORY_USAGE
 #define HB_NO_MT
+#define HB_NO_SIMD
 #define HB_NO_UCD_UNASSIGNED
 #ifndef NDEBUG
 #define NDEBUG
@@ -187,6 +188,7 @@
 
 #ifdef HB_OPTIMIZE_SIZE_MORE
 #define HB_NO_OT_RULESETS_FAST_PATH
+#define HB_NO_SIMD
 #endif
 
 #ifdef HB_MINIMIZE_MEMORY_USAGE

src/hb-ot-layout-common.hh

diff --git a/src/hb-ot-layout-common.hh b/src/hb-ot-layout-common.hh
index 82f8808..691c9f3 100644
--- a/src/hb-ot-layout-common.hh
+++ b/src/hb-ot-layout-common.hh
@@ -38,9 +38,9 @@
 #include "OT/Layout/Common/Coverage.hh"
 #include "OT/Layout/types.hh"
 
-#ifdef __SSE__
+#if defined(__SSE__) && !defined(HB_NO_SIMD)
 #include <xmmintrin.h>
-#elif defined(__ARM_NEON)
+#elif defined(__ARM_NEON) && !defined(HB_NO_SIMD)
 #include <arm_neon.h>
 #endif
 
@@ -3157,11 +3157,11 @@
       unsigned i = 0;
       for (; i + 4 <= count; i += 4)
       {
-#ifdef __SSE__
+#if defined(__SSE__) && !defined(HB_NO_SIMD)
 	__m128 a = _mm_loadu_ps (in + i);
 	__m128 b = _mm_loadu_ps (out + i);
 	_mm_storeu_ps (out + i, _mm_add_ps (a, b));
-#elif defined(__ARM_NEON)
+#elif defined(__ARM_NEON) && !defined(HB_NO_SIMD)
 	float32x4_t a = vld1q_f32 (in + i);
 	float32x4_t b = vld1q_f32 (out + i);
 	vst1q_f32 (out + i, vaddq_f32 (a, b));
@@ -3178,18 +3178,18 @@
     else
     {
       unsigned i = 0;
-#ifdef __SSE__
+#if defined(__SSE__) && !defined(HB_NO_SIMD)
       __m128 s = _mm_set1_ps (scalar);
-#elif defined(__ARM_NEON)
+#elif defined(__ARM_NEON) && !defined(HB_NO_SIMD)
       float32x4_t s = vdupq_n_f32 (scalar);
 #endif
       for (; i + 4 <= count; i += 4)
       {
-#ifdef __SSE__
+#if defined(__SSE__) && !defined(HB_NO_SIMD)
 	__m128 a = _mm_loadu_ps (in + i);
 	__m128 b = _mm_loadu_ps (out + i);
 	_mm_storeu_ps (out + i, _mm_add_ps (_mm_mul_ps (a, s), b));
-#elif defined(__ARM_NEON)
+#elif defined(__ARM_NEON) && !defined(HB_NO_SIMD)
 	float32x4_t a = vld1q_f32 (in + i);
 	float32x4_t b = vld1q_f32 (out + i);
 	vst1q_f32 (out + i, vmlaq_f32 (b, a, s));