Speculative fix for building on INTEGRITY with ARM NEON

Avoid using inline assembler

Task-number: QTBUG-72716
Change-Id: I696efb5a787416eb4fc5ba3a250461aaa9a4afc2
Reviewed-by: Janne Koskinen <janne.p.koskinen@qt.io>

4 files changed, 53 insertions, 64 deletions

diff --git a/mkspecs/common/ghs-integrity-armv8.conf b/mkspecs/common/ghs-integrity-armv8.conf
index ee4ce458de..e454cfd245 100644
--- a/mkspecs/common/ghs-integrity-armv8.conf
+++ b/mkspecs/common/ghs-integrity-armv8.conf
@@ -17,8 +17,8 @@ os_directory = $$(INTEGRITY_DIR)
 isEmpty(os_directory): \
     error("This qmakespec requires $INTEGRITY_DIR to be set")
 
-QMAKE_CC                = cxintarm64 -U__ARM_NEON__ -U__ARM_NEON -bsp $$bsp_name -os_dir $$os_directory -non_shared
-QMAKE_CXX               = cxintarm64 -U__ARM_NEON__ -U__ARM_NEON -bsp $$bsp_name -os_dir $$os_directory -non_shared
+QMAKE_CC                = cxintarm64 -bsp $$bsp_name -os_dir $$os_directory -non_shared
+QMAKE_CXX               = cxintarm64 -bsp $$bsp_name -os_dir $$os_directory -non_shared
 QMAKE_LINK              = $$QMAKE_CXX
 QMAKE_AR                = $$QMAKE_CXX -archive -o
 
diff --git a/src/gui/image/qimage_neon.cpp b/src/gui/image/qimage_neon.cpp
index 57a24edeca..9dbcb11db5 100644
--- a/src/gui/image/qimage_neon.cpp
+++ b/src/gui/image/qimage_neon.cpp
@@ -52,65 +52,41 @@ Q_GUI_EXPORT void QT_FASTCALL qt_convert_rgb888_to_rgb32_neon(quint32 *dst, cons
 
     const quint32 *const end = dst + len;
 
-    // align dst on 64 bits
-    const int offsetToAlignOn8Bytes = (reinterpret_cast<quintptr>(dst) >> 2) & 0x1;
-    for (int i = 0; i < offsetToAlignOn8Bytes; ++i) {
+    // align dst on 128 bits
+    const int offsetToAlignOn16Bytes = (reinterpret_cast<quintptr>(dst) >> 2) & 0x3;
+    for (int i = 0; i < offsetToAlignOn16Bytes; ++i) {
         *dst++ = qRgb(src[0], src[1], src[2]);
         src += 3;
     }
 
-    if ((len - offsetToAlignOn8Bytes) >= 8) {
-        const quint32 *const simdEnd = end - 7;
-#if !defined(Q_PROCESSOR_ARM_64)
-        register uint8x8_t fullVector asm ("d3") = vdup_n_u8(0xff);
-        do {
+    if ((len - offsetToAlignOn16Bytes) >= 16) {
+        const quint32 *const simdEnd = end - 15;
+        uint8x16x4_t dstVector;
 #if Q_BYTE_ORDER == Q_BIG_ENDIAN
-            asm volatile (
-                "vld3.8     { d4, d5, d6 }, [%[SRC]] !\n\t"
-                "vst4.8     { d3, d4, d5, d6 }, [%[DST],:64] !\n\t"
-                : [DST]"+r" (dst), [SRC]"+r" (src)
-                : "w"(fullVector)
-                : "memory", "d4", "d5", "d6"
-            );
+        dstVector.val[0] = vdupq_n_u8(0xff);
 #else
-            asm volatile (
-                "vld3.8     { d0, d1, d2 }, [%[SRC]] !\n\t"
-                "vswp d0, d2\n\t"
-                "vst4.8     { d0, d1, d2, d3 }, [%[DST],:64] !\n\t"
-                : [DST]"+r" (dst), [SRC]"+r" (src)
-                : "w"(fullVector)
-                : "memory", "d0", "d1", "d2"
-            );
+        dstVector.val[3] = vdupq_n_u8(0xff);
 #endif
-        } while (dst < simdEnd);
-#else
-        register uint8x8_t fullVector asm ("v3") = vdup_n_u8(0xff);
         do {
+            uint8x16x3_t srcVector = vld3q_u8(src);
+            src += 3 * 16;
 #if Q_BYTE_ORDER == Q_BIG_ENDIAN
-            asm volatile (
-                "ld3     { v4.8b, v5.8b, v6.8b }, [%[SRC]], #24 \n\t"
-                "st4     { v3.8b, v4.8b, v5.8b, v6.8b }, [%[DST]], #32 \n\t"
-                : [DST]"+r" (dst), [SRC]"+r" (src)
-                : "w"(fullVector)
-                : "memory", "v4", "v5", "v6"
-            );
+            dstVector.val[1] = srcVector.val[0];
+            dstVector.val[2] = srcVector.val[1];
+            dstVector.val[3] = srcVector.val[2];
 #else
-            asm volatile (
-                "ld3     { v0.8b, v1.8b, v2.8b }, [%[SRC]], #24 \n\t"
-                "mov v4.8b, v2.8b\n\t"
-                "mov v2.8b, v0.8b\n\t"
-                "mov v0.8b, v4.8b\n\t"
-                "st4     { v0.8b, v1.8b, v2.8b, v3.8b }, [%[DST]], #32 \n\t"
-                : [DST]"+r" (dst), [SRC]"+r" (src)
-                : "w"(fullVector)
-                : "memory", "v0", "v1", "v2", "v4"
-            );
+            dstVector.val[0] = srcVector.val[2];
+            dstVector.val[1] = srcVector.val[1];
+            dstVector.val[2] = srcVector.val[0];
 #endif
+            vst4q_u8(reinterpret_cast<uint8_t*>(dst), dstVector);
+            dst += 16;
         } while (dst < simdEnd);
-#endif
     }
 
-    while (dst != end) {
+    int i = 0;
+    int length = end - dst;
+    SIMD_EPILOGUE(i, length, 15) {
         *dst++ = qRgb(src[0], src[1], src[2]);
         src += 3;
     }
diff --git a/src/gui/painting/painting.pri b/src/gui/painting/painting.pri
index c3585a4647..a90abed4c0 100644
--- a/src/gui/painting/painting.pri
+++ b/src/gui/painting/painting.pri
@@ -131,9 +131,11 @@ ARCH_HASWELL_SOURCES += painting/qdrawhelper_avx2.cpp
 
 NEON_SOURCES += painting/qdrawhelper_neon.cpp painting/qimagescale_neon.cpp
 NEON_HEADERS += painting/qdrawhelper_neon_p.h
-NEON_ASM += ../3rdparty/pixman/pixman-arm-neon-asm.S painting/qdrawhelper_neon_asm.S
 !uikit:!win32:contains(QT_ARCH, "arm"): CONFIG += no_clang_integrated_as
-!uikit:!win32:!contains(QT_ARCH, "arm64"): DEFINES += ENABLE_PIXMAN_DRAWHELPERS
+!uikit:!win32:!integrity:!contains(QT_ARCH, "arm64") {
+    NEON_ASM += ../3rdparty/pixman/pixman-arm-neon-asm.S painting/qdrawhelper_neon_asm.S
+    DEFINES += ENABLE_PIXMAN_DRAWHELPERS
+}
 
 MIPS_DSP_SOURCES += painting/qdrawhelper_mips_dsp.cpp
 MIPS_DSP_HEADERS += painting/qdrawhelper_mips_dsp_p.h painting/qt_mips_asm_dsp_p.h
diff --git a/src/gui/painting/qdrawhelper_neon.cpp b/src/gui/painting/qdrawhelper_neon.cpp
index 629dfe2358..3fbd651f96 100644
--- a/src/gui/painting/qdrawhelper_neon.cpp
+++ b/src/gui/painting/qdrawhelper_neon.cpp
@@ -50,7 +50,18 @@ QT_BEGIN_NAMESPACE
 void qt_memfill32(quint32 *dest, quint32 value, int count)
 {
     const int epilogueSize = count % 16;
-#if !defined(Q_PROCESSOR_ARM_64)
+#if defined(Q_CC_GHS) || defined(Q_CC_MSVC)
+    // inline assembler free version:
+    if (count >= 16) {
+        quint32 *const neonEnd = dest + count - epilogueSize;
+        const uint32x4_t valueVector1 = vdupq_n_u32(value);
+        const uint32x4x4_t valueVector4 = { valueVector1, valueVector1, valueVector1, valueVector1 };
+        do {
+            vst4q_u32(dest, valueVector4);
+            dest += 16;
+        } while (dest != neonEnd);
+    }
+#elif !defined(Q_PROCESSOR_ARM_64)
     if (count >= 16) {
         quint32 *const neonEnd = dest + count - epilogueSize;
         register uint32x4_t valueVector1 asm ("q0") = vdupq_n_u32(value);
@@ -84,20 +95,20 @@ void qt_memfill32(quint32 *dest, quint32 value, int count)
 
     switch (epilogueSize)
     {
-    case 15:     *dest++ = value;
-    case 14:     *dest++ = value;
-    case 13:     *dest++ = value;
-    case 12:     *dest++ = value;
-    case 11:     *dest++ = value;
-    case 10:     *dest++ = value;
-    case 9:      *dest++ = value;
-    case 8:      *dest++ = value;
-    case 7:      *dest++ = value;
-    case 6:      *dest++ = value;
-    case 5:      *dest++ = value;
-    case 4:      *dest++ = value;
-    case 3:      *dest++ = value;
-    case 2:      *dest++ = value;
+    case 15:     *dest++ = value; Q_FALLTHROUGH();
+    case 14:     *dest++ = value; Q_FALLTHROUGH();
+    case 13:     *dest++ = value; Q_FALLTHROUGH();
+    case 12:     *dest++ = value; Q_FALLTHROUGH();
+    case 11:     *dest++ = value; Q_FALLTHROUGH();
+    case 10:     *dest++ = value; Q_FALLTHROUGH();
+    case 9:      *dest++ = value; Q_FALLTHROUGH();
+    case 8:      *dest++ = value; Q_FALLTHROUGH();
+    case 7:      *dest++ = value; Q_FALLTHROUGH();
+    case 6:      *dest++ = value; Q_FALLTHROUGH();
+    case 5:      *dest++ = value; Q_FALLTHROUGH();
+    case 4:      *dest++ = value; Q_FALLTHROUGH();
+    case 3:      *dest++ = value; Q_FALLTHROUGH();
+    case 2:      *dest++ = value; Q_FALLTHROUGH();
     case 1:      *dest++ = value;
     }
 }