luke
/
verilator
mirror of https://github.com/verilator/verilator.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Internals: Put AstNode flags in padding gap, initialize spare bits 

The bool AstNode flags fit in a padding gap after m_type. This reduces
memory consumption by about 2% on OpenTitan. Initializing the unused
bits in the flags then avoids a read-modify-write in the constructor
(replacing it with a store constant). Overall verilation speed is about
1% faster.
This commit is contained in:
Geza Lore 2021-07-10 08:01:58 +01:00
parent 61e2e55ba5
commit 825d6b87a8
2 changed files with 28 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
src/V3Ast.cpp
View File

@ -74,9 +74,10 @@ AstNode::AstNode(AstType t, FileLine* fl)
m_clonep = nullptr;
m_cloneCnt = 0;
// Attributes
m_didWidth = false;
m_doingWidth = false;
m_protect = true;
m_flags.didWidth = false;
m_flags.doingWidth = false;
m_flags.protect = true;
m_flags.unused = 0; // Initializing this avoids a read-modify-write on construction
m_user1u = VNUser(0);
m_user1Cnt = 0;
m_user2u = VNUser(0);

38
src/V3Ast.h
View File

@ -1334,8 +1334,24 @@ class AstNode VL_NOT_FINAL {
const AstType m_type; // Node sub-type identifier
// ^ ASTNODE_PREFETCH depends on above ordering of members
// padding - 2 extra bytes here after m_type
int m_cloneCnt; // Mark of when userp was set
// AstType is 2 bytes, so we can stick another 6 bytes after it to utilize what would
// otherwise be padding (on a 64-bit system). We stick the attribute flags and the clone
// count here.
struct {
bool didWidth : 1; // Did V3Width computation
bool doingWidth : 1; // Inside V3Width
bool protect : 1; // Protect name if protection is on
uint16_t unused : 13; // Space for more flags here (there must be 16 bits in total)
} m_flags; // Attribute flags
int m_cloneCnt; // Sequence number for when last clone was made
#if defined(__x86_64__) && defined(__gnu_linux__)
// Only assert this on known platforms, as it only affects performance, not correctness
static_assert(sizeof(m_type) + sizeof(m_flags) + sizeof(m_cloneCnt) <= sizeof(void*),
"packing requires padding");
#endif
AstNodeDType* m_dtypep; // Data type of output or assignment (etc)
AstNode* m_headtailp; // When at begin/end of list, the opposite end of the list
@ -1348,12 +1364,6 @@ class AstNode VL_NOT_FINAL {
AstNode* m_clonep; // Pointer to clone of/ source of this module (for *LAST* cloneTree() ONLY)
static int s_cloneCntGbl; // Count of which userp is set
// Attributes
bool m_didWidth : 1; // Did V3Width computation
bool m_doingWidth : 1; // Inside V3Width
bool m_protect : 1; // Protect name if protection is on
// // Space for more bools here
// This member ordering both allows 64 bit alignment and puts associated data together
VNUser m_user1u; // Contains any information the user iteration routine wants
uint32_t m_user1Cnt; // Mark of when userp was set
@ -1512,17 +1522,17 @@ public:
void fileline(FileLine* fl) { m_fileline = fl; }
bool width1() const;
int widthInstrs() const;
void didWidth(bool flag) { m_didWidth = flag; }
bool didWidth() const { return m_didWidth; }
void didWidth(bool flag) { m_flags.didWidth = flag; }
bool didWidth() const { return m_flags.didWidth; }
bool didWidthAndSet() {
if (didWidth()) return true;
didWidth(true);
return false;
}
bool doingWidth() const { return m_doingWidth; }
void doingWidth(bool flag) { m_doingWidth = flag; }
bool protect() const { return m_protect; }
void protect(bool flag) { m_protect = flag; }
bool doingWidth() const { return m_flags.doingWidth; }
void doingWidth(bool flag) { m_flags.doingWidth = flag; }
bool protect() const { return m_flags.protect; }
void protect(bool flag) { m_flags.protect = flag; }
// TODO stomp these width functions out, and call via dtypep() instead
int width() const;