Remove type-punned unions

Type punning even over a union is not legal C++, and also causes the
compilers to produce poorly performing code.

This has already been fixed for the SSE2 code in bilinear sampling but
not for the NEON code.

Change-Id: Id5e184051e0bd78db730d83ef0dda56ac3206e5b
Reviewed-by: Erik Verbruggen <erik.verbruggen@qt.io>

1 files changed, 33 insertions, 30 deletions

diff --git a/src/gui/painting/qdrawhelper.cpp b/src/gui/painting/qdrawhelper.cpp
index f0d0ac0283..fa4470a486 100644
--- a/src/gui/painting/qdrawhelper.cpp
+++ b/src/gui/painting/qdrawhelper.cpp
@@ -2243,45 +2243,48 @@ static const uint * QT_FASTCALL fetchTransformedBilinearARGB32PM(uint *buffer, c
                     const int16x8_t v_disty_ = vshlq_n_s16(v_disty, 4);
                     int32x4_t v_fdx = vdupq_n_s32(fdx*4);
 
-                    ptrdiff_t secondLine = reinterpret_cast<const uint *>(s2) - reinterpret_cast<const uint *>(s1);
-
-                    union Vect_buffer { int32x4_t vect; quint32 i[4]; };
-                    Vect_buffer v_fx;
-
-                    for (int i = 0; i < 4; i++) {
-                        v_fx.i[i] = fx;
-                        fx += fdx;
-                    }
+                    int32x4_t v_fx = vmovq_n_s32(fx);
+                    fx += fdx;
+                    v_fx = vsetq_lane_s32(fx, v_fx, 1);
+                    fx += fdx;
+                    v_fx = vsetq_lane_s32(fx, v_fx, 2);
+                    fx += fdx;
+                    v_fx = vsetq_lane_s32(fx, v_fx, 3);
+                    fx += fdx;
 
                     const int32x4_t v_ffff_mask = vdupq_n_s32(0x0000ffff);
                     const int32x4_t v_fx_r = vdupq_n_s32(0x0800);
 
                     while (b < boundedEnd) {
-
-                        Vect_buffer tl, tr, bl, br;
-
-                        Vect_buffer v_fx_shifted;
-                        v_fx_shifted.vect = vshrq_n_s32(v_fx.vect, 16);
-
-                        int32x4_t v_distx = vshrq_n_s32(vaddq_s32(vandq_s32(v_fx.vect, v_ffff_mask), v_fx_r), 12);
-
-                        for (int i = 0; i < 4; i++) {
-                            int x1 = v_fx_shifted.i[i];
-                            const uint *addr_tl = reinterpret_cast<const uint *>(s1) + x1;
-                            const uint *addr_tr = addr_tl + 1;
-                            tl.i[i] = *addr_tl;
-                            tr.i[i] = *addr_tr;
-                            bl.i[i] = *(addr_tl+secondLine);
-                            br.i[i] = *(addr_tr+secondLine);
-                        }
-
+                        uint32x4x2_t v_top, v_bot;
+
+                        int32x4_t v_fx_shifted = vshrq_n_s32(v_fx, 16);
+
+                        int x1 = vgetq_lane_s32(v_fx_shifted, 0);
+                        v_top = vld2q_lane_u32(s1 + x1, v_top, 0);
+                        v_bot = vld2q_lane_u32(s2 + x1, v_bot, 0);
+                        x1 = vgetq_lane_s32(v_fx_shifted, 1);
+                        v_top = vld2q_lane_u32(s1 + x1, v_top, 1);
+                        v_bot = vld2q_lane_u32(s2 + x1, v_bot, 1);
+                        x1 = vgetq_lane_s32(v_fx_shifted, 2);
+                        v_top = vld2q_lane_u32(s1 + x1, v_top, 2);
+                        v_bot = vld2q_lane_u32(s2 + x1, v_bot, 2);
+                        x1 = vgetq_lane_s32(v_fx_shifted, 3);
+                        v_top = vld2q_lane_u32(s1 + x1, v_top, 3);
+                        v_bot = vld2q_lane_u32(s2 + x1, v_bot, 3);
+
+                        int32x4_t v_distx = vshrq_n_s32(vaddq_s32(vandq_s32(v_fx, v_ffff_mask), v_fx_r), 12);
                         v_distx = vorrq_s32(v_distx, vshlq_n_s32(v_distx, 16));
 
-                        interpolate_4_pixels_16_neon(vreinterpretq_s16_s32(tl.vect), vreinterpretq_s16_s32(tr.vect), vreinterpretq_s16_s32(bl.vect), vreinterpretq_s16_s32(br.vect), vreinterpretq_s16_s32(v_distx), v_disty, v_disty_, colorMask, invColorMask, v_256, b);
+                        interpolate_4_pixels_16_neon(
+                                    vreinterpretq_s16_u32(v_top.val[0]), vreinterpretq_s16_u32(v_top.val[1]),
+                                    vreinterpretq_s16_u32(v_bot.val[0]), vreinterpretq_s16_u32(v_bot.val[1]),
+                                    vreinterpretq_s16_s32(v_distx), v_disty, v_disty_,
+                                    colorMask, invColorMask, v_256, b);
                         b+=4;
-                        v_fx.vect = vaddq_s32(v_fx.vect, v_fdx);
+                        v_fx = vaddq_s32(v_fx, v_fdx);
                     }
-                    fx = v_fx.i[0];
+                    fx = vgetq_lane_s32(v_fx, 0);
 #endif
                 }