Have vvp_vector8_t avoid allocating tiny scalar arrays.(cherry picked from commit 35fe8fae00)

vvp/vvp_net.cc

@@ -1758,12 +1758,12 @@ ostream& operator<< (ostream&out, const vvp_vector2_t&that)
 vvp_vector8_t::vvp_vector8_t(const vvp_vector8_t&that)
 {
       size_ = that.size_;
-      if (size_==0) {
+      if (size_ <= PTR_THRESH) {
-	    bits_ = 0;
+	    memcpy(val_, that.val_, sizeof(val_));
       } else {
-	    bits_ = new vvp_scalar_t[size_];
+	    ptr_ = new vvp_scalar_t[size_];
 	    for (unsigned idx = 0 ;  idx < size_ ;  idx += 1)
-		  bits_[idx] = that.bits_[idx];
+		  ptr_[idx] = that.ptr_[idx];
       }
 }
 
@@ -1771,26 +1771,29 @@ vvp_vector8_t::vvp_vector8_t(const vvp_vector4_t&that,
 			     unsigned str0, unsigned str1)
 : size_(that.size())
 {
-      if (size_ == 0) {
+      if (size_ == 0)
-	    bits_ = 0;
 	    return;
-      }
 
-      bits_ = new vvp_scalar_t[size_];
+      vvp_scalar_t*tmp;
+      if (size_ <= PTR_THRESH)
+	    tmp = new (val_) vvp_scalar_t[PTR_THRESH];
+      else
+	    tmp = ptr_ = new vvp_scalar_t[size_];
+
       for (unsigned idx = 0 ;  idx < size_ ;  idx += 1)
-	    bits_[idx] = vvp_scalar_t (that.value(idx), str0, str1);
+	    tmp[idx] = vvp_scalar_t (that.value(idx), str0, str1);
 
 }
 
 vvp_vector8_t& vvp_vector8_t::operator= (const vvp_vector8_t&that)
 {
 	// Assign to self.
-      if (size_ > 0 && bits_ == that.bits_)
+      if (size_ > PTR_THRESH && that.size_ > PTR_THRESH && ptr_ == that.ptr_)
 	    return *this;
 
       if (size_ != that.size_) {
-	    if (size_ > 0)
+	    if (size_ > PTR_THRESH)
-		  delete[]bits_;
+		  delete[]ptr_;
 	    size_ = 0;
       }
 
@@ -1799,13 +1802,19 @@ vvp_vector8_t& vvp_vector8_t::operator= (const vvp_vector8_t&that)
 	    return *this;
       }
 
+      if (that.size_ <= PTR_THRESH) {
+	    size_ = that.size_;
+	    memcpy(val_, that.val_, sizeof(val_));
+	    return *this;
+      }
+
       if (size_ == 0) {
 	    size_ = that.size_;
-	    bits_ = new vvp_scalar_t[size_];
+	    ptr_ = new vvp_scalar_t[size_];
       }
 
       for (unsigned idx = 0 ;  idx < size_ ;  idx += 1)
-	    bits_[idx] = that.bits_[idx];
+	    ptr_[idx] = that.ptr_[idx];
 
       return *this;
 }
@@ -1814,9 +1823,12 @@ vvp_vector8_t vvp_vector8_t::subvalue(unsigned base, unsigned wid) const
 {
       vvp_vector8_t tmp (wid);
 
+      vvp_scalar_t* tmp_ptr = tmp.size_<=PTR_THRESH? reinterpret_cast<vvp_scalar_t*>(tmp.val_) : tmp.ptr_;
+      const vvp_scalar_t* ptr = size_<=PTR_THRESH? reinterpret_cast<const vvp_scalar_t*>(val_) : ptr_;
+
       unsigned idx = 0;
       while ((idx < wid) && (base+idx < size_)) {
-	    tmp.bits_[idx] = bits_[base+idx];
+	    tmp_ptr[idx] = ptr[base+idx];
 	    idx += 1;
       }
 
@@ -1828,10 +1840,13 @@ vvp_vector8_t part_expand(const vvp_vector8_t&that, unsigned wid, unsigned off)
       assert(off < wid);
       vvp_vector8_t tmp (wid);
 
+      vvp_scalar_t* tmp_ptr = tmp.size_<=vvp_vector8_t::PTR_THRESH? reinterpret_cast<vvp_scalar_t*>(tmp.val_) : tmp.ptr_;
+      const vvp_scalar_t* that_ptr = that.size_<=vvp_vector8_t::PTR_THRESH? reinterpret_cast<const vvp_scalar_t*>(that.val_) : that.ptr_;
+
       unsigned idx = off;
 
       while (idx < wid && that.size_ > (idx-off)) {
-	    tmp.bits_[idx] = that.bits_[idx-off];
+	    tmp_ptr[idx] = that_ptr[idx-off];
 	    idx += 1;
       }
 

vvp/vvp_net.h

@@ -21,6 +21,7 @@
 
 # include  "config.h"
 # include  <stddef.h>
+# include  <new>
 # include  <assert.h>
 
 #ifdef HAVE_IOSFWD
@@ -581,8 +582,15 @@ class vvp_vector8_t {
       vvp_vector8_t& operator= (const vvp_vector8_t&that);
 
     private:
+	// This is the number of vvp_scalar_t objects we can keep in
+	// the val_ buffer. If the vector8 is bigger then this, then
+	// resort to allocations to get a larger buffer.
+      enum { PTR_THRESH = 8 };
       unsigned size_;
-      vvp_scalar_t*bits_;
+      union {
+	    vvp_scalar_t*ptr_;
+	    char val_[PTR_THRESH * sizeof(vvp_scalar_t)];
+      };
 };
 
   /* Resolve uses the default Verilog resolver algorithm to resolve
@@ -612,17 +620,35 @@ extern ostream& operator<< (ostream&, const vvp_vector8_t&);
 inline vvp_vector8_t::vvp_vector8_t(unsigned size)
 : size_(size)
 {
-      if (size_ == 0) {
+      if (size_ <= PTR_THRESH) {
-	    bits_ = 0;
+	    new (val_) vvp_scalar_t[PTR_THRESH];
-	    return;
+      } else {
+	    ptr_ = new vvp_scalar_t[size_];
       }
-      bits_ = new vvp_scalar_t[size_];
 }
 
 inline vvp_vector8_t::~vvp_vector8_t()
 {
-      if (size_ > 0)
+      if (size_ > PTR_THRESH)
-	    delete[]bits_;
+	    delete[]ptr_;
+}
+
+inline vvp_scalar_t vvp_vector8_t::value(unsigned idx) const
+{
+      assert(idx < size_);
+      if (size_ <= PTR_THRESH)
+	    return reinterpret_cast<const vvp_scalar_t*>(val_) [idx];
+      else
+	    return ptr_[idx];
+}
+
+inline void vvp_vector8_t::set_bit(unsigned idx, vvp_scalar_t val)
+{
+      assert(idx < size_);
+      if (size_ <= PTR_THRESH)
+	    reinterpret_cast<vvp_scalar_t*>(val_) [idx] = val;
+      else
+	    ptr_[idx] = val;
 }
 
   // Exactly-equal for vvp_vector8_t is common and should be as tight
@@ -634,26 +660,17 @@ inline bool vvp_vector8_t::eeq(const vvp_vector8_t&that) const
       if (size_ == 0)
 	    return true;
 
+      if (size_ <= PTR_THRESH)
+	    return 0 == memcmp(val_, that.val_, sizeof(val_));
+
       for (unsigned idx = 0 ;  idx < size_ ;  idx += 1) {
-	    if (! bits_[idx] .eeq( that.bits_[idx] ))
+	    if (! ptr_[idx] .eeq( that.ptr_[idx] ))
 		return false;
       }
 
       return true;
 }
 
-inline vvp_scalar_t vvp_vector8_t::value(unsigned idx) const
-{
-      assert(idx < size_);
-      return bits_[idx];
-}
-
-inline void vvp_vector8_t::set_bit(unsigned idx, vvp_scalar_t val)
-{
-      assert(idx < size_);
-      bits_[idx] = val;
-}
-
 /*
  * This class implements a pointer that points to an item within a
  * target. The ptr() method returns a pointer to the vvp_net_t, and