1 files changed, 27 insertions, 14 deletions

diff --git a/cpu_ref/rsCpuScript.cpp b/cpu_ref/rsCpuScript.cpp
index f4abe67f..6c534cf9 100644
--- a/cpu_ref/rsCpuScript.cpp
+++ b/cpu_ref/rsCpuScript.cpp
@@ -258,6 +258,11 @@ static void setCompileArguments(std::vector<const char*>* args,
     args->push_back("-mtriple");
     args->push_back(DEFAULT_TARGET_TRIPLE_STRING);
 
+    // Enable workaround for A53 codegen by default.
+#if defined(__aarch64__) && !defined(DISABLE_A53_WORKAROUND)
+    args->push_back("-aarch64-fix-cortex-a53-835769");
+#endif
+
     // Execute the bcc compiler.
     if (useRSDebugContext) {
         args->push_back("-rs-debug-ctx");
@@ -381,7 +386,6 @@ RsdCpuScriptImpl::RsdCpuScriptImpl(RsdCpuReferenceImpl *ctx, const Script *s) {
     mFieldIsObject = nullptr;
     mForEachSignatures = nullptr;
 #else
-    mCompilerContext = nullptr;
     mCompilerDriver = nullptr;
     mExecutable = nullptr;
 #endif
@@ -409,17 +413,9 @@ bool RsdCpuScriptImpl::init(char const *resName, char const *cacheDir,
 #ifndef RS_COMPATIBILITY_LIB
     bool useRSDebugContext = false;
 
-    mCompilerContext = nullptr;
     mCompilerDriver = nullptr;
     mExecutable = nullptr;
 
-    mCompilerContext = new bcc::BCCContext();
-    if (mCompilerContext == nullptr) {
-        ALOGE("bcc: FAILS to create compiler context (out of memory)");
-        mCtx->unlockMutex();
-        return false;
-    }
-
     mCompilerDriver = new bcc::RSCompilerDriver();
     if (mCompilerDriver == nullptr) {
         ALOGE("bcc: FAILS to create compiler driver (out of memory)");
@@ -1033,14 +1029,34 @@ void RsdCpuScriptImpl::invokeFreeChildren() {
 void RsdCpuScriptImpl::invokeFunction(uint32_t slot, const void *params,
                                       size_t paramLength) {
     //ALOGE("invoke %p %p %i %p %i", dc, script, slot, params, paramLength);
+    void * ap = nullptr;
+
+#if defined(__x86_64__)
+    // The invoked function could have input parameter of vector type for example float4 which
+    // requires void* params to be 16 bytes aligned when using SSE instructions for x86_64 platform.
+    // So try to align void* params before passing them into RS exported function.
+
+    if ((uint8_t)(uint64_t)params & 0x0F) {
+        if ((ap = (void*)memalign(16, paramLength)) != nullptr) {
+            memcpy(ap, params, paramLength);
+        } else {
+            ALOGE("x86_64: invokeFunction memalign error, still use params which is not 16 bytes aligned.");
+        }
+    }
+#endif
 
     RsdCpuScriptImpl * oldTLS = mCtx->setTLS(this);
     reinterpret_cast<void (*)(const void *, uint32_t)>(
 #ifndef RS_COMPATIBILITY_LIB
-        mExecutable->getExportFuncAddrs()[slot])(params, paramLength);
+        mExecutable->getExportFuncAddrs()[slot])(ap ? (const void *)ap : params, paramLength);
 #else
-        mInvokeFunctions[slot])(params, paramLength);
+        mInvokeFunctions[slot])(ap ? (const void *)ap : params, paramLength);
 #endif
+
+#if defined(__x86_64__)
+    if (ap) free(ap);
+#endif
+
     mCtx->setTLS(oldTLS);
 }
 
@@ -1199,9 +1215,6 @@ RsdCpuScriptImpl::~RsdCpuScriptImpl() {
         }
     }
 
-    if (mCompilerContext) {
-        delete mCompilerContext;
-    }
     if (mCompilerDriver) {
         delete mCompilerDriver;
     }