Add VL_CACHE_LINE_BYTES and use 64 as defult.

include/verilatedos.h

@@ -340,6 +340,7 @@ typedef unsigned long long      vluint64_t;     ///< 64-bit unsigned type
 #define VL_QUADSIZE 64                  ///< Bits in a QData / quadword
 #define VL_EDATASIZE 32                 ///< Bits in a EData (WData entry)
 #define VL_EDATASIZE_LOG2 5             ///< log2(VL_EDATASIZE)
+#define VL_CACHE_LINE_BYTES 64          ///< Bytes in a cache line (for alignment)
 
 /// Bytes this number of bits needs (1 bit=1 byte)
 #define VL_BYTES_I(nbits) (((nbits) + (VL_BYTESIZE - 1)) / VL_BYTESIZE)

src/V3EmitC.cpp

@@ -2778,7 +2778,7 @@ void EmitCImp::emitInt(AstNodeModule* modp) {
     }
 
     puts("}");
-    if (!VN_IS(modp, Class)) puts(" VL_ATTR_ALIGNED(128)");
+    if (!VN_IS(modp, Class)) puts(" VL_ATTR_ALIGNED(VL_CACHE_LINE_BYTES)");
     puts(";\n");
 
     emitIntFuncDecls(modp, false);

src/V3EmitCSyms.cpp

@@ -475,7 +475,7 @@ void EmitCSyms::emitSymHdr() {
         puts("void "+protect("__Vdeserialize")+"(VerilatedDeserialize& os);\n");
     }
     puts("\n");
-    puts("} VL_ATTR_ALIGNED(64);\n");
+    puts("} VL_ATTR_ALIGNED(VL_CACHE_LINE_BYTES);\n");
 
     ofp()->putsEndGuard();
 }

src/Verilator.cpp

@@ -466,6 +466,7 @@ static void process() {
         && !v3Global.opt.xmlOnly()
         && !v3Global.opt.dpiHdrOnly()) {
         // Create AstCUse to determine what class forward declarations/#includes needed in C
+        // Must be before V3EmitC
         V3CUse::cUseAll(v3Global.rootp());
 
         // emitcInlines is first, as it may set needHInlines which other emitters read