From 821aa1ff095ae66a89eb8725650dccac363f06ad Mon Sep 17 00:00:00 2001
From: Allan Sandfeld Jensen <allan.jensen@qt.io>
Date: Fri, 24 Jun 2022 14:27:18 +0200
Subject: Avoid SSE2 punning

It is technical UB, even if GCC promises to let it work, but it also
generates inefficient code.

Pick-to: 6.4
Change-Id: I8f0cae3490d32287ecbaa16b1e9ace84223cda2a
Reviewed-by: Thiago Macieira <thiago.macieira@intel.com>
---
 src/gui/painting/qdrawhelper_sse2.cpp | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/src/gui/painting/qdrawhelper_sse2.cpp b/src/gui/painting/qdrawhelper_sse2.cpp
index f26ad1a943..22014ec175 100644
--- a/src/gui/painting/qdrawhelper_sse2.cpp
+++ b/src/gui/painting/qdrawhelper_sse2.cpp
@@ -571,13 +571,14 @@ void qt_scale_image_argb32_on_argb32_sse2(uchar *destPixels, int dbpl,
 
         __m128i srcxVector = _mm_set_epi32(srcx, srcx + ix, srcx + ix + ix, srcx + ix + ix + ix);
 
-        for (; x<w - 3; x += 4) {
-            union Vect_buffer { __m128i vect; quint32 i[4]; };
-            Vect_buffer addr;
-            addr.vect = _mm_srli_epi32(srcxVector, 16);
+        for (; x < (w - 3); x += 4) {
+            const int idx0 = _mm_extract_epi16(srcxVector, 1);
+            const int idx1 = _mm_extract_epi16(srcxVector, 3);
+            const int idx2 = _mm_extract_epi16(srcxVector, 5);
+            const int idx3 = _mm_extract_epi16(srcxVector, 7);
             srcxVector = _mm_add_epi32(srcxVector, ixVector);
 
-            const __m128i srcVector = _mm_set_epi32(src[addr.i[0]], src[addr.i[1]], src[addr.i[2]], src[addr.i[3]]);
+            const __m128i srcVector = _mm_set_epi32(src[idx0], src[idx1], src[idx2], src[idx3]);
             BLEND_SOURCE_OVER_ARGB32_SSE2_helper(dst, srcVector, nullVector, half, one, colorMask, alphaMask);
         }
 
-- 
cgit v1.2.3