From 4f37cf2ce544dad5b207161536566627b6b0b1ef Mon Sep 17 00:00:00 2001
From: Nick Shaforostoff <shaforostoff@gmail.com>
Date: Sun, 28 Feb 2021 20:02:18 +0100
Subject: [PATCH] string 16<->8 bits conversion: SIMD on arm32

this adds emulation for 2 NEON commands that armv7 lacks
this increases conversion speed by around 50% in my simple tests

Change-Id: I4f52d353184e9a8d88089de60e17bd5670637c0c
Reviewed-by: Allan Sandfeld Jensen <allan.jensen@qt.io>
---
 src/corelib/global/qsimd_p.h          | 16 ++++++++++++++++
 src/corelib/text/qstring.cpp          |  4 ++--
 src/corelib/text/qstringconverter.cpp |  4 ++--
 3 files changed, 20 insertions(+), 4 deletions(-)

diff --git a/src/corelib/global/qsimd_p.h b/src/corelib/global/qsimd_p.h
index 2f2d49348f..fbb5f0f021 100644
--- a/src/corelib/global/qsimd_p.h
+++ b/src/corelib/global/qsimd_p.h
@@ -274,6 +274,22 @@ QT_END_NAMESPACE
 // __ARM_NEON__ is not defined on AArch64, but we need it in our NEON detection.
 #define __ARM_NEON__
 #endif
+
+#ifndef Q_PROCESSOR_ARM_64 // vaddv is only available on Aarch64
+inline uint16_t vaddvq_u16(uint16x8_t v8)
+{
+    const uint64x2_t v2 = vpaddlq_u32(vpaddlq_u16(v8));
+    const uint64x1_t v1 = vadd_u64(vget_low_u64(v2), vget_high_u64(v2));
+    return vget_lane_u16(vreinterpret_u16_u64(v1), 0);
+}
+
+inline uint8_t vaddv_u8(uint8x8_t v8)
+{
+    const uint64x1_t v1 = vpaddl_u32(vpaddl_u16(vpaddl_u8(v8)));
+    return vget_lane_u8(vreinterpret_u8_u64(v1), 0);
+}
+#endif
+
 #endif
 // AArch64/ARM64
 #if defined(Q_PROCESSOR_ARM_V8) && defined(__ARM_FEATURE_CRC32)
diff --git a/src/corelib/text/qstring.cpp b/src/corelib/text/qstring.cpp
index 24e725ee9a..49e56d26d6 100644
--- a/src/corelib/text/qstring.cpp
+++ b/src/corelib/text/qstring.cpp
@@ -333,7 +333,7 @@ const char16_t *QtPrivate::qustrchr(QStringView str, char16_t c) noexcept
                                    [=](int i) { return n[i] == c; },
                                    [=](int i) { return n + i; });
 #  endif
-#elif defined(__ARM_NEON__) && defined(Q_PROCESSOR_ARM_64) // vaddv is only available on Aarch64
+#elif defined(__ARM_NEON__)
     const uint16x8_t vmask = { 1, 1 << 1, 1 << 2, 1 << 3, 1 << 4, 1 << 5, 1 << 6, 1 << 7 };
     const uint16x8_t ch_vec = vdupq_n_u16(c);
     for (const char16_t *next = n + 8; next <= e; n = next, next += 8) {
@@ -1004,7 +1004,7 @@ static int ucstrncmp(const QChar *a, const QChar *b, size_t l)
     };
     return UnrollTailLoop<3>::exec(l, 0, lambda, lambda);
 #endif
-#if defined(__ARM_NEON__) && defined(Q_PROCESSOR_ARM_64) // vaddv is only available on Aarch64
+#ifdef __ARM_NEON__
     if (l >= 8) {
         const QChar *end = a + l;
         const uint16x8_t mask = { 1, 1 << 1, 1 << 2, 1 << 3, 1 << 4, 1 << 5, 1 << 6, 1 << 7 };
diff --git a/src/corelib/text/qstringconverter.cpp b/src/corelib/text/qstringconverter.cpp
index 524bb4cc03..df9efe7f67 100644
--- a/src/corelib/text/qstringconverter.cpp
+++ b/src/corelib/text/qstringconverter.cpp
@@ -61,7 +61,7 @@ enum { Endian = 0, Data = 1 };
 static const uchar utf8bom[] = { 0xef, 0xbb, 0xbf };
 
 #if (defined(__SSE2__) && defined(QT_COMPILER_SUPPORTS_SSE2)) \
-    || (defined(__ARM_NEON__) && defined(Q_PROCESSOR_ARM_64))
+    || defined(__ARM_NEON__)
 static Q_ALWAYS_INLINE uint qBitScanReverse(unsigned v) noexcept
 {
 #if defined(__cpp_lib_int_pow2) && __cpp_lib_int_pow2 >= 202002L
@@ -360,7 +360,7 @@ static void simdCompareAscii(const char8_t *&src8, const char8_t *end8, const ch
     src8 += offset;
     src16 += offset;
 }
-#elif defined(__ARM_NEON__) && defined(Q_PROCESSOR_ARM_64) // vaddv is only available on Aarch64
+#elif defined(__ARM_NEON__)
 static inline bool simdEncodeAscii(uchar *&dst, const ushort *&nextAscii, const ushort *&src, const ushort *end)
 {
     uint16x8_t maxAscii = vdupq_n_u16(0x7f);