Vectorize SetTargetAtTime processing in AudioParamTimeline

cdumez · cdumez · commit e212e1fac13a · 2020-09-18T16:43:27.000Z
https://bugs.webkit.org/show_bug.cgi?id=216673 <rdar://problem/69111432> Reviewed by Sam Weinig. Fix the SSE implementation of VectorMath::vsadd() to properly deal with source and destination alignment, similarly to what is done in other VectorMath functions. No new tests, fixes existing crash on GTK bots. * platform/audio/VectorMath.cpp: (WebCore::VectorMath::vsadd): Canonical link: https://commits.webkit.org/229503@main git-svn-id: https://svn.webkit.org/repository/webkit/trunk@267240 268f45cc-cd09-0410-ab3c-d52691b4dbfc
diff --git a/Source/WebCore/ChangeLog b/Source/WebCore/ChangeLog
@@ -1,3 +1,19 @@
+2020-09-18  Chris Dumez  <cdumez@apple.com>
+
+        Vectorize SetTargetAtTime processing in AudioParamTimeline
+        https://bugs.webkit.org/show_bug.cgi?id=216673
+        <rdar://problem/69111432>
+
+        Reviewed by Sam Weinig.
+
+        Fix the SSE implementation of VectorMath::vsadd() to properly deal with source
+        and destination alignment, similarly to what is done in other VectorMath functions.
+
+        No new tests, fixes existing crash on GTK bots.
+
+        * platform/audio/VectorMath.cpp:
+        (WebCore::VectorMath::vsadd):
+
 2020-09-18  Zalan Bujtas  <zalan@apple.com>
 
         [LFC][IFC] Move computedLineLogicalRect to InlineFormattingContext::Geometry
diff --git a/Source/WebCore/platform/audio/VectorMath.cpp b/Source/WebCore/platform/audio/VectorMath.cpp
@@ -264,17 +264,40 @@ void vsadd(const float* sourceP, int sourceStride, const float* scalar, float* d
 
 #if CPU(X86_SSE2)
     if (sourceStride == 1 && destStride == 1) {
-        __m128* pSource;
-        __m128 scalarVector = _mm_set_ps1(*scalar);
+        // If the sourceP address is not 16-byte aligned, the first several frames (at most three) should be processed separately.
+        while ((reinterpret_cast<size_t>(sourceP) & 0x0F) && n) {
+            *destP = *sourceP + *scalar;
+            sourceP++;
+            destP++;
+            n--;
+        }
 
+        // Now the sourceP address is aligned and start to apply SSE.
         int group = n / 4;
-        while (group--) {
-            pSource = reinterpret_cast<__m128*>(const_cast<float*>(sourceP));
-            __m128 dest = _mm_add_ps(*pSource, scalarVector);
-            _mm_storeu_ps(destP, dest);
+        __m128 mScalar = _mm_set_ps1(*scalar);
+        __m128* pSource;
+        __m128* pDest;
+        __m128 dest;
 
-            sourceP += 4;
-            destP += 4;
+        bool destAligned = !(reinterpret_cast<size_t>(destP) & 0x0F);
+        if (destAligned) { // all aligned
+            while (group--) {
+                pSource = reinterpret_cast<__m128*>(const_cast<float*>(sourceP));
+                pDest = reinterpret_cast<__m128*>(destP);
+                *pDest = _mm_add_ps(*pSource, mScalar);
+
+                sourceP += 4;
+                destP += 4;
+            }
+        } else {
+            while (group--) {
+                pSource = reinterpret_cast<__m128*>(const_cast<float*>(sourceP));
+                dest = _mm_add_ps(*pSource, mScalar);
+                _mm_storeu_ps(destP, dest);
+
+                sourceP += 4;
+                destP += 4;
+            }
         }
 
         // Non-SSE handling for remaining frames which is less than 4.
